Alzheimer's disease image feature selection method based on intuitionistic fuzzy encoder

Abstract

The invention provides an Alzheimer's disease image feature selection method based on an intuitive fuzzy encoder, belongs to the technical field of medical information intelligent processing, and solves the technical problems of high-dimensional feature redundancy and poor classification effect during Alzheimer's disease analysis. According to the technical scheme, the method comprises the following steps that S10, a tested Alzheimer's disease test sample set is collected and preprocessed; s20, calculating the membership degree and the non-membership degree of each sample, and constructing a score matrix based on an intuitionistic fuzzy set; s30, pre-training an intuitionistic fuzzy automatic encoder model of a global sample; and S40, constructing a sparse feature coding network, transferring learning parameters, and performing sparse processing on the weight matrix to obtain important features. The method has the advantages that the problem of pathology data feature redundancy can be effectively solved, the data dimension is reduced so as to reduce the calculation requirement, the classification accuracy is improved, and technical support is provided for actual clinical diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent processing of medical information, and in particular to a method for selecting features of Alzheimer's disease images based on an intuitionistic fuzzy encoder. Technical Background

[0002] Alzheimer's disease is a slowly progressive neurodegenerative disease characterized by progressively worsening cognitive impairment, involving functional impairments in memory, language, thinking, judgment, etc. Patients may initially show mild memory loss, and as the disease progresses, they gradually become less able to carry out daily activities and have personality changes, and may eventually completely lose the ability to take care of themselves.

[0003] In the past few years, scholars have developed feature selection methods such as mutual information method, variance selection method and moth flame optimization algorithm to help understand the data, reduce irrelevant and redundant features, and reduce data dimensions to reduce computing requirements. However, the above-mentioned feature selection methods have problems such as sensitivity to noisy data, inability to process linear data and easy to fall into local optimal solutions.

[0004] The deep learning technology that has emerged in recent years has brought new solutions to the field of smart medical technology. Autoencoders can automatically learn and recognize medical images with complex features, have powerful feature extraction and pattern recognition capabilities, and have good prospects in the field of medical image processing. Summary of the invention

[0005] In order to solve the above problems, the present invention provides an Alzheimer's disease image feature selection method based on an intuitionistic fuzzy encoder. In the data processing stage, the intuitionistic fuzzy membership is applied to explore the connection between the data, and different weights are assigned to the data for preprocessing. At the same time, a one-to-one selection layer is introduced to reduce irrelevant and redundant features, reduce the data dimension to reduce the computing requirements, and improve the efficiency of Alzheimer's disease brain image feature selection. The method has a strong application value for Alzheimer's disease brain image feature selection.

[0006] The inventive idea of ​​the present invention is as follows: first, the present invention divides the Alzheimer's disease test sample set; then, the membership and non-membership of each sample are calculated to construct a score matrix based on an intuitionistic fuzzy set; next, an intuitionistic fuzzy encoder model of a global sample is pre-trained; finally, a sparse feature encoding network is constructed, the encoder parameters during pre-training are loaded, and the weight matrix is ​​sparsely processed to obtain important features.

[0007] The present invention is achieved by the following measures: a method for selecting features of Alzheimer's disease brain images based on an intuitionistic fuzzy autoencoder, comprising the following steps:

[0008] S10: Collect samples of Alzheimer's brain images of subjects, build a decision information system, and divide the data into a training sample set and a test sample set with a division ratio of 9:1;

[0009] S20: Calculate the membership and non-membership of each sample and construct a score matrix based on intuitionistic fuzzy sets;

[0010] S30: Use the training sample set to pre-train an intuitionistic fuzzy autoencoder model of a global sample, perform forward propagation and back propagation during the training process, and use the mean square error loss function to update the network parameters;

[0011] S40: Build a sparse feature encoding network, load the encoder parameters during pre-training, and perform sparse processing on the weight matrix to obtain important features.

[0012] Furthermore, the step S10 includes the following steps:

[0013] S11: Collect original samples of magnetic resonance imaging, apply brain imaging data processing and analysis to denoise the data, assign labels to samples according to clinical needs, and establish an original decision information system S = <U, C, D>, where U = {x1, x2..., x i} represents the sample set of Alzheimer's brain image data, x i represents the i-th sample, i=1,2,3,...,740; C={a1,a2,...a j} represents the number of features in the Alzheimer's brain image data, a j represents the jth feature, j = 1, 2, 3, ... 4005; D = {d1, d2, ..., d i} represents the class label of Alzheimer's brain image data, d i represents the class label of the i-th sample, where d i = {0,1}, indicating that the sample does not have / has Alzheimer's disease.

[0014] S12: Divide the above-mentioned Alzheimer's disease brain image data into a training sample set and a test sample set, with a division ratio of 9:1.

[0015] Furthermore, the step S20 includes the following steps:

[0016] S21: Use membership value and non-membership value to bidirectionally judge the fuzzy relationship between samples and categories of Alzheimer's disease brain image dataset and construct intuitionistic fuzzy set A e ={x i ,d i ,R(x i ),eR(x i )}, where R(xi ) represents the membership value of the sample, eR(x i ) represents the non-membership value of the sample;

[0017] S22: Calculate the symptomatic class center C1 and asymptomatic class sample C0:

[0018]

[0019] Among them, N1 is the number of symptomatic samples, and N0 is the number of asymptomatic samples;

[0020] S23: Calculate the symptomatic class radius B1 and the asymptomatic class radius B0 of the sample:

[0021]

[0022] S24: Construct sample x according to the symptomatic and asymptomatic classes of the data set i The membership function is mapped to the space of [0,1] to represent the sample x i The importance of is defined as follows:

[0023]

[0024] Among them, ψ represents the kernel mapping, which maps the data to the high-dimensional feature space, and σ is the bias number;

[0025] S25: Calculate x i For potential heterogeneous samples, construct their non-membership function eR(xi), which is defined as follows:

[0026] eR(x i )=(1-R(x i ))Υ(x i ) (4)

[0027] Υ(xi) represents sample x i The proportion of heterogeneous data points in its neighborhood is expressed as follows:

[0028]

[0029] Among them, ζ is an adjustable parameter for creating a neighborhood;

[0030] S26: Compare and divide the calculated membership function value and non-membership function value, Γ(x i ) represents the quality function of each sample, representing sample x i The fuzzy similarity relationship corresponding to the class center is expressed as follows:

[0031]

[0032] Let the Alzheimer's disease dataset be N is the number of samples, n is the number of features, and the quality matrix is ​​obtained as follows:

[0033] S=diag(Γ(x1),Γ(x2),...,Γ(x 4005 )) (7).

[0034] Furthermore, the step S30 includes the following steps:

[0035] S31: Initialize the parameters of the encoder and decoder θ = {θ enc ,θ dec}, initialize the data set X according to the quality matrix S in step S26 to obtain the initial feature mapping matrix Z = SX;

[0036] S32: Take the current batch Z from the training data set b , use the autoencoder f(Zb,θenc) to transform the batch Z b Mapped to a low-dimensional space, we get the potential representation z of the encoder network b ;

[0037] S33: Use the decoder to receive the latent representation z b , map it back to the original data space, and get the reconstructed original data The decoder g(·) is defined as follows:

[0038] g(z b ,θ dec )=W2·ReLU(W1·z b +b1)+b2 (8)

[0039] Where W1 and W2 represent the weight matrices of the first hidden layer and the output layer, respectively, b1 and b2 are bias terms, and ReLU(·) is the activation function of the hidden layer;

[0040] S34: Use the mean square error loss function to calculate the initial feature mapping matrix Z and the data reconstructed by the decoder The differences between are shown as follows:

[0041]

[0042] in, is the Frobenius norm, which represents the sum of the squared differences of all elements in the matrix;

[0043] S35: Update the parameter θ using the gradient descent method based on the calculated loss value and return the optimized optimal parameter value

[0044] Furthermore, the step S40 includes the following steps:

[0045] S41: Construct a sparse feature encoding network, which includes a one-to-one selection layer and an encoder. The one-to-one selection layer assigns an initial weight w to each feature for feature screening, which is expressed as follows:

[0046] Z′=ZW (10)

[0047] Where W = diag(w), w ≥ 0, the weight w reflects the importance of the feature, and the most relevant features are identified and selected after training;

[0048] S42: Using transfer learning, inherit the best encoder parameters learned in step S35 To update the weight parameters, the objective function of the training process is expressed as follows:

[0049]

[0050] in Represents the pre-trained encoder network The generated latent representation, represents the l1 regularization term of the encoder, and ||w||1 represents the l1 regularization term of the one-to-one selection layer, which is used to sparsify the weight matrix to obtain important features.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] 1. The present invention provides a method for selecting features of Alzheimer's disease brain images based on an intuitionistic fuzzy autoencoder, which introduces intuitionistic fuzzy membership, mines the connections between data, and assigns different weights to the data for preprocessing. This method improves the accuracy and reliability of disease classification.

[0053] 2. The method of pre-training an intuitive fuzzy autoencoder model of a global sample enables the model to learn a more general feature expression form and has better adaptability to the data distribution of unseen samples. This method improves the reconstruction and classification performance of the migration model.

[0054] 3. The present invention constructs a sparse feature selection network method, which introduces a one-to-one selection layer on the basis of ensuring the reconstruction effect, reduces irrelevant and redundant features, reduces data dimensions to reduce computing requirements. This method improves the efficiency of feature selection of Alzheimer's brain images. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The technical solutions and beneficial effects of the present invention will be made apparent by describing in detail the specific embodiments of the present invention in conjunction with the accompanying drawings.

[0056] Figure 1 The present invention is a flowchart of a method for selecting features of Alzheimer's disease brain images based on an intuitionistic fuzzy autoencoder in an embodiment of the present invention.

[0057] Figure 2 4 is a block diagram of a method for selecting features of Alzheimer's disease brain images based on an intuitionistic fuzzy autoencoder in an embodiment of the present invention.

[0058] Figure 3 This is a pre-trained global sample model diagram of an Alzheimer's disease brain image feature selection method based on an intuitionistic fuzzy autoencoder in an embodiment of the present invention.

[0059] Figure 4 This is a diagram of a sparse feature coding network for screening important features of a method for selecting features of Alzheimer's disease brain images based on an intuitionistic fuzzy autoencoder in an embodiment of the present invention. DETAILED DESCRIPTION

[0060] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0061] Example 1

[0062] This embodiment provides a breast cancer data classification method based on multi-granularity evidence neighborhood rough sets. Figures 1 to 4 As shown, the following steps are included:

[0063] S10: Collect samples of Alzheimer's brain images of subjects, build a decision information system, and divide the data into a training sample set and a test sample set with a division ratio of 9:1;

[0064] S20: Calculate the membership and non-membership of each sample and construct a score matrix based on intuitionistic fuzzy sets;

[0065] S30: Use the training sample set to pre-train an intuitionistic fuzzy autoencoder model of a global sample, perform forward propagation and back propagation during the training process, and use the mean square error loss function to update the network parameters;

[0066] S40: Build a sparse feature encoding network, load the encoder parameters during pre-training, and perform sparse processing on the weight matrix to obtain important features.

[0067] Specifically, step S10 includes the following steps:

[0068] S11: Collect original samples of magnetic resonance imaging, apply brain imaging data processing and analysis to denoise the data, assign labels to samples according to clinical needs, and establish an original decision information system S = <U, C, D>, where U = {x1, x2..., x i} represents the sample set of Alzheimer's brain image data, x i represents the i-th sample, i=1,2,3,...,740; C={a1,a2,...a j} represents the number of features in the Alzheimer's brain image data, a j represents the jth feature, j = 1, 2, 3, ... 4005; D = {d1, d2, ..., d i} represents the class label of Alzheimer's brain image data, d i represents the class label of the i-th sample, where d i ={0,1}, indicating that the sample does not have / has Alzheimer's disease. The original data set is converted into a decision information system as shown in Table 1:

[0069] Table 1

[0070]

[0071] S12: Divide the above Alzheimer's disease brain image data into a training sample set and a test sample set, with a division ratio of 9:1. The sample set database, the number of training samples, and the number of test samples are:

[0072]

[0073] Specifically, step S20 includes the following steps:

[0074] S21: Use membership value and non-membership value to judge the fuzzy relationship between data set samples and categories, and construct the intuitive fuzzy set A e ={x i ,d i ,R(x i ),eR(x i )}, where R(x i ) represents the membership value of the sample, eR(x i ) represents the non-membership value of the sample;

[0075] S22: Calculate the symptomatic class center C1 and asymptomatic class sample C0:

[0076]

[0077] Among them, N1 is the number of symptomatic samples, and N0 is the number of asymptomatic samples;

[0078] S23: Calculate the symptomatic class radius B1 and the asymptomatic class radius B0 of the sample:

[0079]

[0080] S24: Construct sample x according to the symptomatic and asymptomatic classes of the data set i The membership function is mapped to the space of [0,1] to represent the sample x i The importance of is defined as follows:

[0081]

[0082] Among them, ψ represents the kernel mapping, which maps the data to the high-dimensional feature space, and σ is the bias number;

[0083] S25: Calculate x i For potential heterogeneous samples, construct their non-membership function eR(xi), which is defined as follows:

[0084] eR(x i )=(1-R(x i ))Υ(x i ) (4)

[0085] Υ(xi) represents sample x i The proportion of heterogeneous data points in its neighborhood is expressed as follows:

[0086]

[0087] Among them, ζ is an adjustable parameter for creating a neighborhood;

[0088] S26: Compare and divide the calculated membership function value and non-membership function value, Γ(x i ) represents the quality function of each sample, representing sample x i The fuzzy similarity relationship corresponding to the class center is expressed as follows:

[0089]

[0090] Table 2 is the quality function for calculating each sample, which can be used to obtain the fuzzy similarity relationship between samples:

[0091] Table 2

[0092]

[0093] Let the Alzheimer's disease dataset be N is the number of samples, n is the number of features, and the quality matrix is ​​obtained, which is specifically expressed as follows:

[0094] S=diag(Γ(x1),Γ(x2),...,Γ(x4005 )) (7).

[0095] Specifically, step S30 includes the following steps:

[0096] S31: Initialize the parameters of the encoder and decoder θ = {θ enc ,θ dec}, initialize the data set X according to the quality matrix S in step S26, and obtain the initial feature mapping matrix Z = SX. Substituting the calculated results into the following:

[0097]

[0098] S32: Take the current batch Z from the training data set b , use the autoencoder f(Zb,θenc) to transform the batch Z b Mapped to a low-dimensional space, we get the potential representation z of the encoder network b ;

[0099] S33: Use the decoder to receive the latent representation z b , map it back to the original data space, and get the reconstructed original data The decoder g(·) is defined as follows:

[0100] g(z b ,θ dec )=W2·ReLU(W1·z b +b1)+b2 (8)

[0101] Where W1 and W2 represent the weight matrices of the first hidden layer and the output layer, respectively, b1 and b2 are bias terms, and ReLU(·) is the activation function of the hidden layer;

[0102] S34: Use the mean square error loss function to calculate the initial feature mapping matrix Z and the data reconstructed by the decoder The differences between are shown as follows:

[0103]

[0104] in, is the Frobenius norm, which represents the sum of the squared differences of all elements in the matrix;

[0105] S35: Update the parameter θ using the gradient descent method based on the calculated loss value and return the optimized optimal parameter value

[0106] Specifically, the step S40 includes the following steps:

[0107] S41: Construct a sparse feature encoding network, including a selection layer and an encoder. The selection layer assigns an initial weight w to each feature for feature screening, which is expressed as follows:

[0108] Z′=ZW (10)

[0109] Where W = diag(w), w ≥ 0, the weight w reflects the importance of the feature, and the most relevant features are identified and selected after training;

[0110] S42: Using transfer learning, inherit the best encoder parameters learned in step S35 To update the weight parameters, the objective function of the training process is expressed as follows:

[0111]

[0112] in Represents the pre-trained encoder network The generated latent representation, represents the l1 regularization term of the encoder, ||w||1 represents the l1 regularization term of the one-to-one selection layer, which is used to sparse the weight matrix to obtain important features. In this example, the batch size is set to 20, the number of global encoder iterations is 400, and the number of feature selection encoder iterations is 600. The Alzheimer's disease brain image feature selection method based on intuitionistic fuzzy autoencoder finally selects the important feature subset [1,2,3,5,7,9,…,3998,4001,4002,4003].

[0113] Example 2

[0114] Referring to Example 1, this example will use the parameters and results calculated in Example 1 to compare with the traditional method to prove the superiority of this example. In the specific comparison, we used different models and compared in different indicators. The final results show that this example is preferred over the traditional method.

[0115] 1. Traditional Model

[0116] CNN model: When the CNN (Convolutional Neural Network) model is used for the Alzheimer's disease feature selection task, the brain image data is first preprocessed, including denoising, standardization, and segmentation. Then, CNN processes the image layer by layer through multiple convolutional layers and pooling layers to automatically extract local features related to the disease, such as changes in the thickness of the cerebral cortex and atrophy of the hippocampus. Disadvantages: The CNN network is deep and has high computational complexity. When processing large-scale image data, the computing resource requirements and storage requirements are relatively large, resulting in longer training time and higher hardware resource consumption.

[0117] DNN model: The DNN (Deep Neural Network) model normalizes and denoises the Alzheimer's brain images through preprocessing steps and then passes them into multiple fully connected layers. DNN learns high-level abstract features of input data through multiple layers of nonlinear transformations. The output of each layer is processed by an activation function to gradually extract complex features related to Alzheimer's brain images. Disadvantages: DNNs generally lack the ability to perceive spatial structures and cannot effectively capture local structural features in images like other neural networks.

[0118] 2. Comparison indicators

[0119] This example is used to perform feature selection tasks on Alzheimer's brain images, using classification accuracy (ACC) as the evaluation indicator. We use three different classification models: support vector machine (SVM), logistic regression (LR), and multi-layer perceptron (MLP) for evaluation.

[0120] 3. Classifier Model

[0121] Support vector machine: The core idea is to find a hyperplane in the feature space so that samples of different categories are separated as much as possible while maximizing the interval between categories. Support vector machines perform well in high-dimensional spaces and are particularly suitable for data sets with dimensions higher than the number of samples.

[0122] Logistic regression: It is suitable for binary classification problems and can quickly find the optimal separating hyperplane to achieve higher classification accuracy. Logistic regression is robust and avoids the risk of overfitting.

[0123] Multilayer Perceptron: By simulating the working mode of biological neural system, multiple fully connected layers are used to perform feature learning and classification decisions. Each layer of the multilayer perceptron has a nonlinear activation function, which can process data with complex nonlinear relationships, surpassing the traditional linear classifier.

[0124] 4. Comparison results

[0125] As can be seen from the chart, the present embodiment performs well in the evaluation indicators and is significantly better than other models. On the three classifiers of support vector machine (SVM), logistic regression (LR), and multi-layer perceptron (MLP), the feature subset classification effect after feature selection in the present embodiment reaches 77.03% in the accuracy of support vector machine (SVM), which is significantly higher than 62.82% of DNN and 66.67% of CNN, indicating that the feature selection effect of the present method in high-dimensional feature data sets is better. The accuracy of the present method in logistic regression (LR) reaches 81.08%, which is significantly higher than 65.38% of DNN and 67.95% of CNN, indicating that the training effect of the present embodiment is more robust and has high robustness. The classification accuracy of the present method in multi-layer perceptron (MLP) reaches 83.78%, which is significantly higher than 65.38% of DNN and 67.95% of CNN, indicating that the present embodiment can learn the nonlinear relationship between data and process more complex data. The feature selection model of this embodiment can select important features to accurately classify Alzheimer's disease in different classifiers, showing its superior performance. The specific comparison is shown in Table 3:

[0126] Table 3

[0127]

[0128] The above description is only an exemplary embodiment of the present invention, and does not limit the scope of patent protection of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the scope of patent protection of the present invention.

Claims

1. A method for selecting features of Alzheimer's disease images based on an intuitionistic fuzzy encoder, characterized in that: The following steps are involved: S10: Collect samples of Alzheimer's brain images of subjects, build a decision information system, and divide the data into a training sample set and a test sample set with a division ratio of 9:1; S20: Calculate the membership and non-membership of each sample and construct a score matrix based on intuitionistic fuzzy sets; S30: Use the training sample set to pre-train an intuitionistic fuzzy autoencoder model of a global sample, perform forward propagation and back propagation during the training process, and use the mean square error loss function to update the network parameters; S40: Build a sparse feature encoding network, load the encoder parameters during pre-training, and perform sparse processing on the weight matrix to obtain important features.

2. The Alzheimer's disease image feature selection method based on intuitionistic fuzzy encoder according to claim 1 is characterized in that: The step S10 comprises the following steps: S11: Collect original samples of magnetic resonance imaging, apply brain imaging data processing and analysis to denoise the data, assign labels to samples according to clinical needs, and establish an original decision information system S = <U, C, D>, where U = {x1, x2..., x i } represents the sample set of Alzheimer's brain image data, x i represents the i-th sample, i=1,2,3,...,740; C={a1,a2,...a j } represents the number of features in the Alzheimer's brain image data, a j represents the jth feature, j = 1, 2, 3, ... 4005; D = {d1, d2, ..., d i } represents the class label of Alzheimer's brain image data, d i represents the class label of the i-th sample, where d i ={0,1}, indicating that the sample does not have / has Alzheimer's disease; S12: Divide the above-mentioned Alzheimer's disease brain image data into a training sample set and a test sample set, with a division ratio of 9:

1.

3. The method for selecting features of Alzheimer's disease images based on an intuitionistic fuzzy encoder according to claim 1, characterized in that: The step S20 comprises the following steps: S21: Use membership value and non-membership value to bidirectionally judge the fuzzy relationship between samples and categories of Alzheimer's disease brain image dataset and construct intuitionistic fuzzy set A e ={x i ,d i ,R(x i ),eR(x i )}, where R(x i ) represents the membership value of the sample, eR(x i ) represents the non-membership value of the sample; S22: Calculate the symptomatic class center C1 and asymptomatic class sample C0: Among them, N1 is the number of symptomatic samples, and N0 is the number of asymptomatic samples; S23: Calculate the symptomatic class radius B1 and the asymptomatic class radius B0 of the sample: S24: Construct sample x according to the symptomatic and asymptomatic classes of the data set i The membership function is mapped to the space of [0,1] to represent the sample x i The importance of is defined as follows: Among them, ψ represents the kernel mapping, which maps the data to the high-dimensional feature space, and σ is the bias number; S25: Calculate x i For potential heterogeneous samples, construct their non-membership function eR(xi), which is defined as follows: eR(x i )=(1-R(x i ))Υ(x i ) (4) Υ(xi) represents sample x i The proportion of heterogeneous data points in its neighborhood is expressed as follows: Among them, ζ is an adjustable parameter for creating a neighborhood; S26: Compare and divide the calculated membership function value and non-membership function value, Γ(x i ) represents the quality function of each sample, representing sample x i The fuzzy similarity relationship corresponding to the class center is expressed as follows: Let the Alzheimer's disease dataset be N is the number of samples, n is the number of features, and the quality matrix is ​​obtained as follows: S=diag(Γ(x1),Γ(x2),...,Γ(x 4005 )) (7).

4. The Alzheimer's disease image feature selection method based on intuitionistic fuzzy encoder according to claim 1, characterized in that: The step S30 comprises the following steps: S31: Initialize the parameters of the encoder and decoder θ = {θ enc ,θ dec }, initialize the data set X according to the quality matrix S in step S26 to obtain the initial feature mapping matrix Z = SX; S32: Take the current batch Z from the training data set b , use the autoencoder f(Zb,θenc) to transform the batch Z b Mapped to a low-dimensional space, we get the potential representation z of the encoder network b ; S33: Use the decoder to receive the latent representation z b , map it back to the original data space, and get the reconstructed original data The decoder g(·) is defined as follows: g(z b ,θ dec )=W2 ReLU(W1 with b +b1)+b2 (8) Where W1 and W2 represent the weight matrices of the first hidden layer and the output layer, respectively, b1 and b2 are bias terms, and ReLU(·) is the activation function of the hidden layer; S34: Use the mean square error loss function to calculate the initial feature mapping matrix Z and the data reconstructed by the decoder The differences between are shown as follows: in, is the Frobenius norm, which represents the sum of the squared differences of all elements in the matrix; S35: Update the parameter θ using the gradient descent method based on the calculated loss value and return the optimized optimal parameter value 5. The method for selecting Alzheimer's disease image features based on intuitionistic fuzzy encoder according to claim 1, characterized in that: The step S40 comprises the following steps: S41: Construct a sparse feature encoding network, which includes a one-to-one selection layer and an encoder. The one-to-one selection layer assigns an initial weight w to each feature for feature screening, which is expressed as follows: Z′=ZW (10) Where W = diag(w), w ≥ 0, the weight w reflects the importance of the feature, and the most relevant features are identified and selected after training; S42: Using transfer learning, inherit the best encoder parameters learned in step S35 To update the weight parameters, the objective function of the training process is expressed as follows: in Represents the pre-trained encoder network The generated latent representation, represents the l1 regularization term of the encoder, and ||w||1 represents the l1 regularization term of the selection layer, which is used to sparse the weight matrix to obtain important features.

Images