Medical image recognition method for enhancing fuzzy hyperbox neural network

Abstract

The invention provides a medical image recognition method for enhancing a fuzzy hyperbox neural network. The enhanced fuzzy super-box neural network is widely applied to the field of image recognition due to the online learning ability of the enhanced fuzzy super-box neural network, and the method improves a fuzzy super-box neural network model and applies the improved fuzzy super-box neural network model to medical image recognition. The model firstly adopts a gray level co-occurrence matrix to carry out texture feature extraction on a breast cancer self-organizing pathology image, key feature information is provided for subsequent image recognition, secondly, extracted feature data is input into an enhanced fuzzy super-box neural network model, a space search optimization algorithm is adopted to optimize a generated super-box, and a new super-box neural network is obtained. And finally, outputting a recognition result that the breast cancer self-organizing is the pathological image.

Description

Technical Field

[0001] The invention relates to the field of medical image recognition, and in particular to a method for enhancing a fuzzy hyperbox neural network. Background Art

[0002] Medical image recognition is one of the important applications of artificial intelligence in the medical field. It helps doctors diagnose diseases more quickly and accurately through the analysis and processing of medical images, thereby providing patients with timely and effective treatment plans. In recent years, with the rapid development of artificial intelligence technologies such as deep learning and computer vision, the field of medical image recognition has made rapid progress. However, medical images in the real world are highly diverse and complex. When a new data set appears, the model may not be able to adapt well to these new features and changes, because the new data set may contain situations that the model has never encountered before or new pathological manifestations. This requires the model to be retrained, resulting in a waste of resources. Fuzzy hyperbox neural networks are widely used in the field of medical images due to their online learning capabilities, but traditional fuzzy hyperbox neural network models have some limitations. One of the main limitations is that the hyperboxes between different categories have fuzzy hyperbox boundaries, which means that data between different categories will be confused during recognition, resulting in a decrease in the recognition ability of the model. Summary of the invention

[0003] The purpose of the present invention is to overcome the above-mentioned technical shortcomings and deficiencies and to provide a medical image recognition method based on an enhanced fuzzy hyperbox neural network.

[0004] The technical solution adopted by the present invention is as follows:

[0005] A medical image recognition method based on enhanced fuzzy hyperbox neural network comprises the following steps:

[0006] Step 1: Obtain a medical image dataset and use the gray-level co-occurrence matrix to extract image texture features. The experimental data uses the BreakHis dataset, a public medical image database. In the experiment, the images in the database are first converted to grayscale. Then, for these grayscale images, a gray-level co-occurrence matrix is ​​constructed at a distance of 1 pixel in the 0-degree direction, and 13 key texture features are extracted from the matrix.

[0007] Step 2: Construct an enhanced fuzzy hyperbox neural network and use information gain theory to select important features to avoid the influence of invalid features on hyperbox construction. The hyperbox construction part cancels the hyperbox shrinkage operation that causes data distortion in the original hyperbox construction and uses an optimization algorithm to optimize the hyperbox.

[0008] Step 3: Enhance the fuzzy hyperbox neural network optimization and use the space search optimization algorithm to optimize the hyperbox, so as to avoid the overlap problem between hyperboxes of different categories. The hyperbox generated in step 2 is used as the initial hyperbox, and a population is generated near the initial hyperbox. The population is updated and iterated through global space search, local space search and opposing space search.

[0009] Step 4: Output the medical image recognition results, identify the input pattern, and output the medical image classification results.

[0010] Step 1: Obtain a medical image dataset and use the gray-level co-occurrence matrix to extract image texture features. The specific steps are as follows:

[0011] Step 1.1: The experiment uses the BreakHis dataset of breast cancer tissue pathology images, which consists of 7909 breast tumor tissue microscopic images collected from 82 patients at different magnifications (40x, 100x, 200x, and 400x). It contains 2480 benign samples and 5429 malignant samples;

[0012] Step 1.2: Preprocess the image and convert the color image into a grayscale image;

[0013] Step 1.3: Determine the adjacent direction and distance, the adjacent direction is 0° and the distance is 1;

[0014] Step 1.4: Construct the gray-level co-occurrence matrix P(i,j), where the size of the matrix is ​​N×N, where N is the number of gray levels of the image;

[0015] Step 1.5, normalize the gray level co-occurrence matrix;

[0016] Step 1.6: Calculate Haralick texture features, including energy, contrast, correlation, variance, etc.:

[0017] Energy calculation:

[0018]

[0019] Where P(i,j) is an element in the gray-level co-occurrence matrix, which represents the probability of gray level i and gray level j co-occurring;

[0020] Contrast calculation, the contrast reflects the degree of difference in gray levels in the image, that is, the amplitude of the change in pixel values ​​in the image:

[0021]

[0022] Correlation calculation, correlation is used to measure the correlation of gray-level co-occurrence matrix elements in the row or column direction, where μ x μ y is the mean, σx σ y is the standard deviation:

[0023]

[0024] Entropy calculation, entropy is used to measure the complexity and variation of texture in an image:

[0025]

[0026] The inverse difference matrix calculation reflects the clarity and regularity of the texture:

[0027]

[0028] Variance calculation describes the degree of dispersion of the elements in the matrix, where u is the mean:

[0029]

[0030] And average calculation, and average are used to describe the contrast and brightness of image texture:

[0031]

[0032] The sum variance is used to describe the information of contrast and uniformity in the image texture:

[0033]

[0034] And entropy calculation, and entropy reflects the grayscale and complexity in the image:

[0035]

[0036] Difference variance calculation, which measures the variance of the difference between two gray levels in an image:

[0037]

[0038] Difference entropy calculation, which measures the randomness of the difference distribution between gray levels in the image texture:

[0039]

[0040] Information metric correlation calculation, information metric correlation is a measure used to describe the correlation between image grayscale values:

[0041]

[0042] Calculation of the maximum correlation metric, which is used to evaluate the linear dependence between pixel gray levels in an image:

[0043]

[0044] Among them, σ xy is the covariance:

[0045]

[0046] E x , E y is the expected value:

[0047]

[0048] P x (i), P y (j) are the marginal probabilities of rows and columns.

[0049] Step 2: Construct an enhanced fuzzy hyperbox neural network. The specific steps are as follows:

[0050] Step 2.1: Use information gain theory to select important features:

[0051] H(D|A)=∑(D i | / |D)*H(D i )

[0052]

[0053] Among them, p represents the probability of each category in the data set, D i is the sample subset of the i-th value of feature A, |D i |It is D i size;

[0054] Step 2.2: Build the initial hyperbox:

[0055] Determine whether there is a hyperbox currently: If there is no hyperbox, create a new hyperbox according to the input mode, otherwise determine whether there is a hyperbox of the same category in the current hyperbox set. If there is no hyperbox of the same category, create a new hyperbox, otherwise calculate the membership value of the input mode to the hyperbox of the same category:

[0056]

[0057] Among them, b j (A h ) is the membership function, A h =(a h1 ,a h2 ,...,a hn )∈I n is the hth input function, γ is the sensitivity parameter, V j =(v 1 ,v 2 ,...,vn ) and W j =(w 1 ,w 2 ,...,w n ) represents the maximum and minimum points of the hyperbox;

[0058] Determine whether the winning super box meets the super box expansion conditions:

[0059]

[0060] If it is satisfied, the hyperbox is expanded, otherwise a new hyperbox is created. The expanded hyperbox is as follows:

[0061]

[0062] Step 3: Use the spatial search optimization algorithm to optimize the enhanced fuzzy hyperbox neural network model. The specific steps are as follows:

[0063] Step 3.1, generate the initial population: the generated hyperbox set is used as the first particle in the population, and then other particles in the initial population are generated by randomly floating up and down in a small range based on the particle;

[0064] Step 3.2: Calculate the pattern classification accuracy and use the accuracy as the fitness function to evaluate the individuals in the population;

[0065] Step 3.3: Global space search, select the solution X with the best fitness value from the r solution spaces best , in X best Generate a new solution F based on 1 :

[0066] F 1 ={X best |X 1i =rand(x min ,x max )}

[0067] Step 3.4: Local space search: randomly select some individuals from the r solution spaces to generate the second solution space F 2 :

[0068]

[0069] Step 3.5: Opposition space search, select the initial r solutions and generate opposition solutions;

[0070] Step 3.6: Select the best r solutions from the generated 2r+2 solutions:

[0071]

[0072] Step 3.7: Determine whether the termination condition is met: If the termination condition is not met, continue to execute steps 3.2 to 3.4, otherwise, output the final fuzzy hyperbox set.

[0073] Step 4: Output medical image recognition results: According to the constructed model, the input pattern is recognized and the classification results are output.

[0074] Advantages and beneficial effects of the present invention

[0075] The present invention proposes a medical image recognition method of an enhanced fuzzy hyperbox neural network, which is an online learning method, so that the model does not need to be retrained when facing new samples. Secondly, in the enhanced fuzzy hyperbox neural network, a spatial search optimization algorithm is used to optimize the hyperbox, avoiding the problem of fuzzy hyperbox boundaries between different categories. In the BreakHis dataset of self-organizing pathological images, the model shows good recognition performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 It is the overall flow chart of the implementation of enhanced fuzzy hyperbox neural network for medical image recognition;

[0077] Figure 2 It is the overall flow chart of the enhanced fuzzy hyperbox neural network model; DETAILED DESCRIPTION

[0078] The present invention is described in detail below with reference to the accompanying drawings and examples.

[0079] The experimental data uses the BreakHis dataset, which was established by the P&D Pathological Anatomy and Cytopathology Laboratory in Paraná, Brazil. The dataset consists of 7,909 breast tumor tissue microscopic images collected from 82 patients at different magnifications (40x, 100x, 200x, and 400x), including 2,480 benign samples and 5,429 malignant samples. Figure 1 The specific implementation scheme shown includes the following contents and steps:

[0080] Step 1: Obtain a medical image dataset and use the gray-level co-occurrence matrix to extract image texture features:

[0081] Step 1.1: Randomly select 1,000 images from each magnification database, of which 500 are benign images and 500 are malignant images. For each magnification dataset, 70% of the data is used for training and the remaining 30% is used for testing.

[0082] Step 1.2: Convert the image from a color image to a grayscale image;

[0083] Step 1.3: Determine the adjacent direction and distance, the adjacent direction is 0° and the distance is 1;

[0084] Step 1.4: Construct the gray-level co-occurrence matrix P(i,j), where the size of the matrix is ​​256×256;

[0085] Step 1.5: Use the gray-level co-occurrence matrix to extract texture features from the image, which includes 13 key texture features including correlation and entropy. At this time, the training set is a 700×13 matrix and the test set is a 300×13 matrix.

[0086] Step 2: Construct an enhanced fuzzy hyperbox neural network model. The enhanced fuzzy hyperbox neural network model is as follows: Figure 2 As shown:

[0087] Step 2.1: Use information gain theory to select m important features, where m is set to two-thirds of the number of features;

[0088] Step 2.2: Construct the initial hyperbox;

[0089] Determine whether there is a hyperbox currently: If there is no hyperbox, create a new hyperbox according to the input mode, otherwise determine whether there is a hyperbox of the same category in the current hyperbox set. If there is no hyperbox of the same category, create a new hyperbox, otherwise calculate the membership value of the input mode to the hyperbox of the same category:

[0090]

[0091] Among them, b j (A h ) is the membership function, A h =(a h1 ,a h2 ,...,a hn )∈I n is the hth input function, γ is the sensitivity parameter, V j =(v 1 ,v 2 ,...,v n ) and W j =(w 1 ,w 2 ,...,w n ) represents the maximum and minimum points of the hyperbox;

[0092] Determine whether the winning super box meets the super box expansion conditions:

[0093]

[0094] If it is satisfied, the hyperbox is expanded, otherwise a new hyperbox is created. The expanded hyperbox is as follows:

[0095]

[0096] Step 3: Use the spatial search optimization algorithm to optimize the generated initial hyperbox:

[0097] Step 3.1, initialize the population: the population size is set to 100, the generated hyperbox is used as the first particle, and the remaining particles are generated by floating up and down within a range of ±0.02 based on this particle;

[0098] Step 3.2, calculate the pattern classification accuracy: use the accuracy as the fitness function to evaluate the individuals in the population;

[0099] Step 3.3: Global space search: Select the solution X with the best fitness value from the 100 solution spaces best , in X best Generate a new solution F based on 1 :

[0100] F 1 ={X best |X 1i =rand(x min ,x max )}

[0101] Step 3.4: Local space search: Randomly select some individuals from the 100 solution spaces to generate the second solution space F 2 :

[0102]

[0103] Step 3.5: Opponent space search: Select the initial 100 solutions and generate the opposing solutions

[0104] Step 3.6: Select the best 100 solutions from the generated 202 solutions:

[0105]

[0106] Step 3.7: Determine whether the termination condition is met: If the termination condition is not met, continue to execute steps 3.2 to 3.4, otherwise, output the final fuzzy hyperbox set.

[0107] Step 4: Output the medical image recognition results and use the constructed enhanced fuzzy hyperbox classification model to train and identify the medical image data.

[0108] The enhanced fuzzy hyperbox neural network model proposed in the present invention was tested 8 times using the BreakHis database. The average results of the 8 experiments were taken, and the average accuracy and average number of hyperboxes were used as evaluation indicators. The hyperbox expansion parameter was set to 0.05. The experimental results are shown in Table 1.

[0109] Table 1 Comparison of BreakHis data classification results

[0110]

[0111] The above description is only a preferred example of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the scope of the claims of the present invention.

Claims

1. A medical image recognition method based on an enhanced fuzzy hyperbox neural network, comprising the following steps: Step 1: Obtain a medical image dataset and use the gray-level co-occurrence matrix to extract image texture features. The experimental data uses the BreakHis dataset, a public medical image database. In the experiment, the images in the database are first converted to grayscale. Then, for these grayscale images, a gray-level co-occurrence matrix is ​​constructed at a distance of 1 pixel in the 0-degree direction, and 13 key texture features are extracted from the matrix. Step 2: Construct an enhanced fuzzy hyperbox neural network and use information gain theory to select important features to avoid the influence of invalid features on hyperbox construction. The hyperbox construction part cancels the hyperbox shrinkage operation that causes data distortion in the original hyperbox construction and uses an optimization algorithm to optimize the hyperbox. Step 3: Enhance the fuzzy hyperbox neural network optimization and use the space search optimization algorithm to optimize the hyperbox, so as to avoid the overlap problem between hyperboxes of different categories. The hyperbox generated in step 2 is used as the initial hyperbox, and a population is generated near the initial hyperbox. The population is updated and iterated through global space search, local space search and opposing space search. Step 4: Output the medical image recognition results, identify the input pattern, and output the medical image classification results.

2. The medical image recognition method of enhanced fuzzy hyperbox neural network according to claim 1 is characterized in that: The steps of obtaining the medical image dataset and extracting the image texture features using the gray-level co-occurrence matrix described in step 1 are as follows: Step 1.1: The dataset used in this experiment is the BreakHis dataset, which contains 7909 images, divided into two types: malignant and benign. They are collected at four different magnifications, namely 40x, 100x, 200x, and 400x. Step 1.2: Convert the image from a color image to a grayscale image; Step 1.3: Determine the adjacent direction and distance, the adjacent direction is 0° and the distance is 1; Step 1.4: Construct the gray-level co-occurrence matrix P(i,j), where the size of the matrix is ​​N×N, where N is the number of gray levels of the image; Step 1.5, normalize the gray level co-occurrence matrix; Step 1.6: Calculate the haralick texture features, including energy, contrast, correlation, entropy, inverse difference matrix, variance, sum average, sum variance, sum entropy, difference variance, difference entropy, information metric correlation and maximum correlation metric 13 textures feature: energy calculate: Where P(i,j) is an element in the gray-level co-occurrence matrix, which represents the probability of gray level i and gray level j co-occurring; Contrast calculation, the contrast reflects the degree of difference in gray levels in the image, that is, the amplitude of the change in pixel values ​​in the image: Correlation calculation, correlation is used to measure the correlation of gray-level co-occurrence matrix elements in the row or column direction, where μ x μ y is the mean, σ x σ y is the standard deviation: Entropy calculation, entropy is used to measure the complexity and variation of texture in an image: The inverse difference matrix calculation reflects the clarity and regularity of the texture: Variance calculation describes the degree of dispersion of the elements in the matrix, where u is the mean: And average calculation, and average are used to describe the contrast and brightness of image texture: The sum variance is used to describe the information of contrast and uniformity in the image texture: And entropy calculation, and entropy reflects the grayscale and complexity in the image: Difference variance calculation, which measures the variance of the difference between two gray levels in an image: Difference entropy calculation, which measures the randomness of the difference distribution between gray levels in the image texture: Information metric correlation calculation, information metric correlation is a measure used to describe the correlation between image grayscale values: Calculation of the maximum correlation metric, which is used to evaluate the linear dependence between pixel gray levels in an image: Among them, σ xy is the covariance: E x , E y is the expected value: P x (i), P y (j) are the marginal probabilities of rows and columns.

3. The medical image recognition method of enhanced fuzzy hyperbox neural network according to claim 1, characterized in that: The construction of the enhanced fuzzy hyperbox neural network described in step 2 uses the information gain theory to select important features, thereby avoiding the influence of invalid features on the hyperbox construction. The hyperbox construction part cancels the hyperbox shrinkage operation that causes data distortion in the original hyperbox construction and uses the optimization algorithm to optimize the hyperbox. The specific steps are as follows: Step 2.1: Use information gain theory to select m important features, where m is set to two-thirds: H(D|A)=∑(D i | / |D)*H(D i ) Among them, p represents the probability of each category in the data set, D i is the sample subset of the i-th value of feature A, |D i |It is D i size; Step 2.2: Build the initial hyperbox: Determine whether there is a hyperbox currently: If there is no hyperbox, create a new hyperbox based on the input pattern; if there is a hyperbox, further check whether there is a hyperbox of the same category in the hyperbox set; if there is no hyperbox of the same category, add a hyperbox; if there is a hyperbox of the same category, calculate the membership of the input pattern relative to these hyperboxes of the same category: Among them, b j (A h ) is the membership function, A h =(a h1 ,a h2 ,...,a hn )∈I n is the hth input function, γ is the sensitivity parameter, V j =(v1,v2,...,v n ) and W j =(w1,w2,...,w n ) represents the maximum and minimum points of the hyperbox; Determine whether the current winning super box meets the super box expansion conditions: Where θ∈[0,1] is the hyperbox expansion coefficient, which defines the size of the hyperbox. If the expansion is satisfied, the hyperbox is expanded, otherwise a new hyperbox is created. The expanded hyperbox is as follows: in Represents the minimum and maximum points of the expanded hyperbox.

4. The medical image recognition method of enhanced fuzzy hyperbox neural network according to claim 1, characterized in that: The enhanced fuzzy hyperbox neural network optimization described in step 3 has the following specific steps: Step 3.1: Generate the initial population. The initial population size is set to 100. The generated hyperbox set is used as the first particle in the population. Then, based on this particle, a small range of random floating up and down is performed to generate other particles in the initial population. Step 3.2: Global space search, select the solution X with the best fitness value from the solution space best , in X best Generate a new solution F1 based on: F1={X best |X 1i =rand(x min ,x max )} where x min , x max are the minimum and maximum values ​​of the solution respectively; Step 3.3: Local space search: randomly select some individuals from the 100 solution spaces to generate the second solution space F2: Where c is the number of individuals selected; Step 3.4: Opposition space search, select the initial 100 solutions and generate opposition solutions; Step 3.5: Select the best 100 solutions from the generated 202 solutions; Step 3.6: Determine whether the termination condition is met: If the termination condition is not met, continue to execute steps 3.2 to 3.4, otherwise, output the final fuzzy hyperbox set.

5. The medical image recognition method of enhanced fuzzy hyperbox neural network according to claim 1, characterized in that: Step 4 outputs the medical image recognition results, identifies the input pattern, and outputs the medical image classification results.

Images