KR20050043869A - Developing a computer aided diagnostic system on breast cancer using adaptive neuro-fuzzy inference system - Google Patents

Abstract

The present invention relates to an algorithm for automating the diagnosis of a disease, using an adaptive fuzzy-neural network or other artificial intelligence algorithm, which is one of artificial intelligence algorithms, for a disease that is difficult to diagnose and does not have high accuracy of diagnosis by a doctor. The present invention relates to an apparatus for enabling automatic diagnosis using computer software or hardware.

In the embodiment of the present invention, by limiting the input variables to excellent factors through a preliminary data processing operation through a genetic algorithm, and by establishing an automatic diagnosis system for breast cancer through an adaptive fuzzy-neural network, the fast execution time and possible accurate results At the same time, we have completed the method to satisfy.

Description

Developing a Computer Aided Diagnostic System on Breast Cancer Using Adaptive Neuro-Fuzzy Inference System

The present invention relates to a system for automatically diagnosing breast cancer and other difficult diseases using a computer, and in particular, using an adaptive neuro-fuzzy inference system (ANFIS), which is one of artificial intelligence algorithms, The present invention relates to a breast cancer automatic diagnosis device for modeling a diagnosis environment of breast cancer and determining whether the disease is infected.

Currently, diagnosis of the disease is based on the results obtained through various tests, but since the final judgment is made entirely by doctors who are experts, there is a possibility of misdiagnosis, and furthermore, However, even when it is difficult to determine the exact relationship between them, even a specialist can be difficult to diagnose.

Breast cancer is known to be the most common tumor-related disease among Korean women, and the incidence rate is continuously increasing due to the westernization of lifestyle and eating habits, and the number of patients / suspected patients is rapidly increasing. There are two main methods used to diagnose benign and malignant breast cancer. Firstly, the breast of a patient suspected of ultrasonography or mammography is photographed using ultrasound or X-ray of breast cancer, and then diagnosed based on the shape, and the accuracy of diagnosis is not as high as 70%. The second method, called the Fine Needle Aspiration Test (FNA test), is a more accurate method and uses a fine needle to extract tissue suspected of breast cancer and make a diagnosis based on the analysis result. In this method, the diagnosis by a conventional doctor has an average of 90% accuracy, and is the preferred method when the breast is not excised and examined.

In fine needle aspiration biopsy, (1) tissue thickness, (2) cell size uniformity, (3) cell shape uniformity, (4) adhesion degree, (5) single epithelial cell size, (6) exposed nucleus nucleus Nine factors such as (7) chromatin, (8) normal cell nucleus, and (9) indirect fission are obtained as a result of the test. The final diagnosis is a combination of these factors and the patient's medical history. By a doctor). In this case, however, these nine elements are not closely related to each other, and because of the strong nonlinearity between them, it is not easy to make an accurate diagnosis, so even an expert may be confused in obtaining a final diagnosis. There is a possibility.

Similarly, in the diagnosis using mammograms using X-ray, ultrasound, etc., there are various forms in the results, and accurate diagnosis is not easy, and as in the former method, skilled and experienced to obtain accurate final diagnosis. Many experts have the disadvantages that are necessary, and moreover, even the diagnosis made by the experts may not always be accurate.

Accordingly, an object of the present invention, which was devised to solve the problems of the prior art operating as described above, is a final diagnostic determination process of breast cancer using results obtained by breast cancer examination through fine needle aspiration test or X-ray, ultrasound, etc. , Using the AI algorithm's learning and reasoning functions, to obtain a final diagnosis, or to obtain a decision that can be referred to in the final diagnosis, with the help of computer software or hardware for a large and difficult to determine test result. .

Another object of the present invention is to assist the computer with the role of a highly skilled specialist required in the existing diagnostic process, thereby reducing the cost required for manpower supply and also freeing the manpower shortage.

In addition, the algorithm used in the present invention can be used to determine the final diagnosis or process of a disease that is not limited to breast cancer and is difficult to obtain another accurate diagnosis. It is also easy to extend and apply to automated diagnostic systems.

Embodiment of the present invention created to achieve the above object,

Processing the data obtained through the fine needle aspiration test or the captured image according to the artificial intelligence system;

Designing an automated diagnostics system using artificial intelligence algorithms such as neural networks, fuzzy systems, and other neuro-fuzzy mixed systems;

Using a genetic algorithm to find factors favorable for diagnosis;

Modeling an automated diagnostic system using an adaptive neuro-fuzzy inference system and existing diagnostic data;

Combining genetic algorithms and adaptive fuzzy-neural networks to find essential and beneficial factors for the diagnostic system and to train the system;

Setting parameters required for genetic algorithms, fuzzy systems, neural networks (neural network circuits);

Obtaining a result when a new input is entered, using a system where learning from the existing data is completed;

Using the test data, testing the system;

Using the finally configured system, obtaining an output;

Etc. are included.

Hereinafter, the operating principle of the present invention will be described in detail with reference to the accompanying drawings. In the following description of the present invention, if it is determined that a detailed description of a known function or configuration may unnecessarily flow the gist of the present invention, the detailed description thereof will be omitted. Terms to be described later are terms defined in consideration of functions in the present invention, and may be changed according to intentions or customs of users or operators. Therefore, the definition should be made based on the contents throughout the specification.

The present invention relates to a system for learning data about breast cancer using an adaptive fuzzy-neural network, which is an artificial intelligence algorithm, and based on the learned model, for a new patient to diagnose whether or not breast cancer is present. Prior to this process, as a pretreatment process, for minimizing the model and the efficiency of the system, among the many input variables derived from fine-grained or mammograms, the most efficient and diagnostic-affected factors were selected. It is found using an algorithm. Therefore, the two algorithms that play a major role in the present invention, the genetic algorithm and the adaptive neuro-fuzzy network will be briefly described.

Genetic algorithms attempt to solve complex real-world problems by simulating on the computer a natural problem-solving method in nature, where genes of an individual evolve to find superior traits over generations. In particular, genetic algorithms are widely applied to adaptive search, learning and optimization problems.

The general problem solving method of genetic algorithm based on this natural evolution is as follows.

First, determine gene populations with different characteristics for any method to be solved, and form the first generation of different gene populations.

Second, individuals within a generation are ranked from the year best suited to the problem they are targeting, based on a set of objective functions determined by the user, and a number of lower ranking entities are eliminated.

Third, by using the genes of surviving individuals, genetic recombination and reproduction occurs, and in this process, operations such as crosses and mutations occur to increase the diversity of the individuals, thereby preventing them from falling into local seas.

Fourth, using the newly constructed generation again, the competition process is repeated, evaluated, and as generations continue, only the better individuals remain in the population, and finally the optimal solution can be obtained.

In the present invention, genetic algorithms have been used to reduce the computational time of heavy computers. The information obtained from fine needle aspiration biopsy or tomography findings of breast cancer varies considerably and can only be achieved if they are integrated and applied to an adaptive flow-purge system. However, there are factors that are less relevant to the results, and the problem is that there is no clear linear relationship between these information or there is no way to identify the information and simply remove unnecessary elements. Therefore, in the present invention, prior to modeling the data, genetic algorithms are used to select input elements that are highly suitable and advantageous for obtaining accurate results in advance, thereby increasing the efficiency of data processing, reducing the complexity of the program, The computation time and high accuracy are obtained.

In this specification, the description by the result of the fine needle aspiration test is demonstrated in various breast cancer test methods. If you want to model the data by learning the adaptive fuzzy-neural network or other learnable AI algorithms using the nine previously described elements obtained by fine needle aspiration biopsy, the problem is that a large amount of sample data The high amount of computation that is caused and the complexity of the finished system. The most influential operation of any AI network is the number of input variables, the number of hidden layer nodes, and the number of iterations of learning.

In the present invention, using the above nine elements and the University of Wisconsin University Breast Cancer Database consisting of 380 sample data, the data was modeled, and an automatic diagnosis system was constructed. Among the inputtable elements, excellent elements were extracted. The steps are as follows and shown in FIG.

First, the nine input elements are reconstructed into binary strings;

Second, to constitute the first generation to be used as the initial input of the genetic algorithm;

Third, construct the objective function of the genetic algorithm using a simplified adaptive neuro-fuzzy system;

Fourth, set parameters related to genetic recombination, crosses, and mutations;

Fifth, using the dominant factors obtained from the genetic algorithm, remodel the adaptive fuzzy-neural network and complete the final diagnostic system.

In this way, it is possible to obtain six input elements having a high contribution and a high correlation among the nine input elements, in which the result and the relationship contribute to obtaining the final diagnosis result. 2 shows the result that the overall error decreases with generations. These results are then used to model the adaptive fuzzy-neural network. The following is an explanation of an adaptive fuzzy-neural network.

System modeling based on typical mathematical methods is not suitable for uncertain and complex systems. On the other hand, fuzzy inference system adopting fuzzy set theory, fuzzy rules and fuzzy inference can express human knowledge quantitatively and can be expressed by linguistic control rules. The adaptive fuzzy-neural network system that extends such fuzzy inference system to the neural network concept and adds a learning algorithm is one of the neuro-fuzzy systems for fuzzy modeling and control system. Adaptive Fuzzy-Nural Network Nodes in each layer perform independent behavior of fuzzy inference system and identify first and second parameters of fuzzy-neural network using hybrid back-propagation algorithm and hybrid learning algorithm using least-squares method.

The structure of a two input one output fuzzy model with two rules is shown in FIG. The latter part of the rule is linear linear, and there are two membership functions assigned to each input variable.

An example of an adaptive neuro-fuzzy network structure using two membership functions for each input variable with two input and one output equivalent to the fuzzy model shown in FIG. 3 is shown in FIG.

Where input is x, y and output is z. The fuzzy rules included here have two rules of linear inference.

The calculation performed at each node of each layer is as follows.

1st floor: The output by each node i is shown in [Equation 2].

here, Is a node Input, Is a fuzzy language variable.

Is Enter x as the membership function value Indicates the degree of satisfaction. Uses the bell-shaped membership function represented by Equation 3.

here, Is a set of parameters. By adjusting these parameters, you change the shape of the membership function. This is the first half parameter to be identified.

Layer 2: Each node multiplies the incoming signals and outputs them as shown in Equation 4. This means calculating the first half fit of each fuzzy rule.

Layer 3: Each node i computes the ratio of the sum of the goodness of fit of all the rules to the goodness of fit of the i th rule as shown in Equation 5. In other words, the output of each node is normalized goodness of fit.

Fourth Floor: Multiply the latter variable for each node.

Where is the output of the third floor and is the later set of parameters to be tuned.

5th floor: Add all input signals and calculate the output of the entire fuzzy model as shown in [Equation 7].

In order to identify each parameter of adaptive fuzzy-neural network, error backpropagation algorithm and least square method based on gradient descent method are used. In this hybrid learning algorithm, the output signal of each node in the omni-directional path is calculated to the fourth floor, and the latter half variable is identified by the least square method while the first half variable is fixed. In the backward path, the error rate is passed back from the output node to the input node and the first half variable is adjusted by the gradient descent method. This hybrid algorithm approach is much faster than the gradient descent approach. In addition, since the first half and second half parameters are separately identified in the identification of the first half and second half parameters, the number of parameters to be identified can be reduced. 5 summarizes each delivery direction.

The error backpropagation algorithm for identifying the first half variable identifies the first half parameter of the fuzzy rule based on the error as follows.

here, Is the learning rate and is calculated as shown in [Equation 9].

New parameters Is given by

From the adaptive fuzzy-neural network shown in FIG. 4, it was found that when the value of the first half parameter is fixed, the overall output value can be expressed as a linear combination of second half parameters. The hybrid learning algorithm shown in the adaptive fuzzy-neural network can be summarized to derive a hybrid learning rule combining gradient descent and least-square method for fast tuning of parameters. This will be described in detail below.

The omni-directional path uses the least squares method and the input vector of the model. With respect to, the following equation holds.

Therefore, Equation 11 is expressed as follows when applied to the entire data model.

In short,

if, If there is an inverse, then the following relationship holds.

If [Equation 13] is changed due to random noise or modeling error, it can be expressed as in [Equation 15].

Here, instead of obtaining the exact parameters for [Equation 3], the square error is minimized. To obtain.

[Equation 5] is When the square error is the smallest. The regular equation that satisfies this is as follows.

Of Equation 17 above New data pairs If is authorized

Is expressed as:

Here, for simplicity Say,

here, Because of,

Wow If you use,

Since Equation 21 can be obtained.

Substituting Equation 21 into Equation 20, When applied to, Equation 22 is obtained.

Thereby silver And new data pairs Is represented here, again

The relationship is established and the final equation is summarized as follows.

The backward path learning uses gradient descent, and the algorithm is as follows.

First, the bell-shaped membership function used is shown in Equation 25.

Here, each parameter to be tuned Therefore, it can be seen that Equation 26 is obtained by taking partial derivatives for each parameter.

Through the above-described process, the adaptive fuzzy-neural network identifies each parameter, learns the system, and models the system. In the embodiment of the present invention, six types of input data obtained in the preceding data processing process using a genetic algorithm are used. By using this, the system is modeled and a final automated diagnostic model is obtained. By inputting test result data from a new patient into the final model thus obtained, a diagnosis result of the patient can be obtained. The overall block diagram is shown in FIG. 6, the diagnosis results obtained from the automatic diagnosis system are shown in FIG. 7, and a diagram of the performance of the configured system is shown in FIG. 8.

6 is a block diagram showing the configuration of an entire system according to a preferred embodiment of the present invention. At the input of the whole system, first, a fine needle aspiration test or imaging, in order to obtain the data necessary and effective in obtaining the output using the genetic algorithm before inputting the data into the adaptive fuzzy-neuro network, the modeling algorithm, as a preprocessor. Variables representing each type of the various types of data obtained by the projection method are defined as binary strings. Then each of these variables is automatically assigned a value of 0 and 1. In this case, 0 means that this input variable is not used as an input of the fuzzy-neuro network, and 1 means that it is used. Subsequent operations are the same as those of the genetic algorithm, and the genetic algorithm is executed using a reduced adaptive fuzzy-neural network as a decision function, generating a new generation through crosses, mutations. As a final result of this, individuals with optimal conditions will be selected, which, in turn, will be used as the final input of a complete fuzzy-neural network to model the diagnosis of breast cancer.

7 is a diagram illustrating test results by an automatic diagnosis system constructed according to a preferred embodiment of the present invention. The graph at the top of FIG. 7 is the result of training the adaptive fuzzy-neural network using the finally obtained input combination, and the graph at the bottom is the result of the test using the learned model. The data marked with 'x' is the exact diagnosis of actual breast cancer data and the data marked with 'o' is the automatic diagnosis obtained by the output of the adaptive fuzzy-neural network. At this time, whether malignant or benign was determined based on 0.5. Based on the results of FIG. 7, the artificial breast cancer diagnosis system presented in the present invention was about 98% in the training data and about 96% in the test data. Has diagnostic accuracy.

8 illustrates the performance of a system constructed in accordance with a preferred embodiment of the present invention. As can be seen in Figure 8, by the genetic algorithm, the diagnostic model modeled using the six inputs obtained has an accuracy of about 97% in terms of accuracy, it shows a difference of about 1% from the model using the full data In contrast, this model is useful because the learning process is completed in about 4 minutes, as opposed to the model using the entire data at the execution time of about 38 minutes. Considering the results of the diagnosis, considering that the accuracy of diagnosis obtained by the doctors who are the actual expert group is about 90%, the accuracy of the diagnosis system presented in the present invention is also very high.

By using the algorithm used in the present invention as described above, it is possible not only to breast cancer, but also to receive the help of the final diagnosis of the disease or the process of the disease that is difficult to obtain another accurate diagnosis. It is also easy to extend and apply to automated diagnostic systems.

Meanwhile, in the detailed description of the present invention, specific embodiments have been described, but various modifications are possible without departing from the scope of the present invention. Therefore, the scope of the present invention should not be limited to the described embodiments, but should be defined not only by the scope of the following claims, but also by those equivalent to the scope of the claims.

In the present invention operating as described in detail above, the effects obtained by the representative ones of the disclosed inventions will be briefly described as follows.

The present invention, in the diagnosis by the data of the high non-linearity breast cancer test that is not easy to make accurate diagnosis even in the expert group of doctors, using the artificial intelligence algorithm adaptive fuzzy-neuron network, the reference during the final diagnosis or doctor diagnosis It is effective to build a system that can be utilized.

In addition, by using the method used in the invention, it can be usefully used in the diagnosis of other diseases by a large amount of data or uncorrelated data.

In addition, the disease can be diagnosed by the method as proposed in the present invention using not only an adaptive fuzzy-neural network but also other AI algorithms.

1 is a step of extracting a high contribution factor among input factors using a genetic algorithm;

2 is a view that the error is reduced with generations in the genetic algorithm;

3 is a structure of a two input one output fuzzy model with two rules;

4 is an adaptive neuro-fuzzy network architecture using two membership functions;

5 is a table illustrating a learning method in an adaptive fuzzy-neural network;

6 is a block diagram of the configuration of the entire system;

7 is a diagram of test results by a configured automated diagnostic system;

8 is a diagram of the performance of a configured system;

Claims (1)

In the diagnosis of breast cancer using an artificial intelligence algorithm, Processing the data obtained through the fine needle aspiration test or the captured image according to the artificial intelligence system; Designing an automated diagnostics system using artificial intelligence algorithms such as neural networks, fuzzy systems, and other neuro-fuzzy mixed systems; Using a genetic algorithm to find factors favorable for diagnosis; Modeling an automated diagnostic system using an adaptive neuro-fuzzy inference system and existing diagnostic data; Combining genetic algorithms and adaptive fuzzy-neural networks to find essential and beneficial factors for the diagnostic system and to train the system; Setting parameters required for genetic algorithms, fuzzy systems, neural networks (neural network circuits); Obtaining a result when a new input is entered, using a system in which learning has been completed from existing data; Using the test data, testing the system; Using the finally configured system, obtaining an output;