A classification method for Alzheimer's disease enhanced by collaborative learning method

Abstract

The present invention discloses a classification method for Alzheimer's disease enhanced based on a collaborative learning method. The method uses Resnet18 as a backbone network, and modifies the single fully connected layer of Resnet18 into no less than two fully connected layers, which are respectively used for Alzheimer's disease and natural image classification; at the same time, the batch normalization layer at all positions of Resnet18 is modified from one to at least two, and finally an attention mechanism module is added to the residual block structure in Resnet18, so as to construct an Alzheimer's disease classification model. Then the classification model is trained by a training set, and the optimal model for the classification of Alzheimer's disease is retained. Finally, the Alzheimer's disease image data to be classified is also normalized in the same way as step two, and then input into the optimal model for the classification of Alzheimer's disease, and the model outputs the corresponding classification result. The present invention introduces collaborative learning training across data sets to enhance the diagnosis of Alzheimer's disease, and the model has strong generalization ability and high robustness.

Description

Technical Field

[0001] The present invention relates to the field of computer medical image analysis, and in particular to a classification method for Alzheimer's disease enhanced by a collaborative learning method. Background Art

[0002] Alzheimer's disease is a chronic neurodegenerative disease, which is often manifested clinically as cognitive impairment, memory impairment, behavioral disorders, and a gradual decline in daily living ability. Studies have shown that AD will slowly destroy the patient's thinking and memory ability, eventually causing the patient to lose the ability to take care of themselves and even die. The use of artificial intelligence technology to assist in the diagnosis of AD is of great significance for intervening in the disease and slowing down the progression of the disease, and can enable family members to take more targeted treatment and care for the disease.

[0003] In current clinical diagnosis, structural magnetic resonance imaging (sMRI) is mainly used to clinically evaluate patients' brain diseases. sMRI can show doctors the patient's brain structure and depict the spatial structural changes of brain tissue. As a non-destructive, non-invasive and low-cost imaging method, sMRI has been widely used in clinical diagnosis.

[0004] In recent years, due to the widespread application of Convolutional Neural Network (CNN) in the field of computer vision, many researchers have applied CNN to the field of medical imaging.

[0005] Traditional methods generally use convolutional neural networks to learn features from slice images of brain images, and send the features to the fully connected layer for classification. This method can extract features from images and classify diseases. The current mainstream method is to use transfer learning technology to extract features from the model and then use the SVM method to classify diseases. The above methods can learn image features to a certain extent and use features for classification, but due to the small amount of medical image data and unclear image features, the performance of the model is often difficult to improve. On the other hand, the above methods do not use an effective attention module to capture differential features of images with unclear features. Summary of the invention

[0006] In order to solve the problem that the generalization ability of the model is not strong under the training of unbalanced data for Alzheimer's disease, this paper proposes a classification method for Alzheimer's disease enhanced by collaborative learning method. The classification model adopts the training method of natural image and medical image cross-dataset, supplements the single change feature extraction of medical image with the features of natural image, and tries to integrate the attention mechanism module CBAM to improve the generalization ability of the model. Finally, the slice voting method is used for the prediction of the model to improve the accuracy of the model prediction.

[0007] The objective of the present invention is achieved through the following technical solutions:

[0008] A classification method for Alzheimer's disease enhanced by collaborative learning method, characterized by comprising the following steps:

[0009] Step 1: Construct a classification model for Alzheimer's disease, which uses Resnet18 as the backbone network and modifies the single fully connected layer of Resnet18 into at least two fully connected layers, which are used for Alzheimer's disease and natural image classification respectively; at the same time, the batch normalization layer at all positions of Resnet18 is modified from one to at least two, and finally an attention mechanism module is added to the residual block structure in Resnet18; the number of the fully connected layers is equal to the number of normalization layers at each position;

[0010] Step 2: Select Alzheimer's disease image data and natural images, and normalize them to construct a training set for Alzheimer's disease and a training set for natural images respectively;

[0011] Step 3: Alternately input the Alzheimer's disease training set and the natural image training set constructed in step 2 into the Alzheimer's disease classification model constructed in step 1, select the corresponding normalization layer and fully connected layer according to the different input training sets, and collaboratively train the model parameters. During the training process, sum up the losses of all training sets and perform back propagation. The predicted probabilities of multiple brain image slices are averaged by the slice voting method, and the average is used as the final output of the model to retain the optimal model for the classification of Alzheimer's disease;

[0012] Step 4: The Alzheimer's disease image data to be classified is also normalized in the same way as in step 2, and then input into the optimal model for Alzheimer's disease classification, and the model outputs the corresponding classification results.

[0013] Furthermore, in step 3, the final output features of the model will select different fully connected layers according to different input data sets to obtain the final classification results:

[0014] y i (x i )=classify i (x i )

[0015] Among them, x i is the data of the i-th data set, classify i To select different classifiers according to different data sets, different classification results y are finally obtained. i .

[0016] Furthermore, in step 3, according to the input training set D i The normalization layers corresponding to different choices are specifically expressed as follows:

[0017]

[0018]

[0019] Among them, BN i Represents the i-th normalization layer at a certain position of Resnet18, represents the jth input of the i-th data set, μ and ε are the mean and variance of the input tensor calculated along the channel direction, and ε in the formula is a constant; After standardization According to D i Different, choose different BN i , that is, different γ i , β i .

[0020] Furthermore, in step 4, during model training, different training sets D are used alternately. i Select a dataset to construct a batch of samples and feed them into the model to obtain the forward propagation loss. When the loss of each dataset through the entire classification model is obtained, the loss is summed and back-propagated.

[0021]

[0022] Where n is the data set D i , classes represents the number of categories, y represents the category variable, y is 0 or 1, if the category is the same as the sample category, y is 1, otherwise y is 0; p represents the predicted probability that the sample belongs to category k after softmax calculation; the model with the smallest Alzheimer's disease loss is selected as the optimal model for the classification of Alzheimer's disease.

[0023] Furthermore, in step three, during the back propagation process, the SGD optimizer is used to optimize the model parameters.

[0024] The beneficial effects of the present invention are as follows:

[0025] The present invention uses a new collaborative learning and training method to alternately perform cross-dataset collaborative training on Alzheimer's disease image slices and natural image datasets, thereby enhancing the generalization ability of the Resnet network and being able to more strongly tap the structural feature extraction capabilities of the convolutional neural network. By adding an attention module to the model, the model can effectively improve the feature information of sensitive areas in the image, allowing the model to more focusedly extract rich and high-quality feature information, and can steadily improve the classification accuracy of the disease. At the same time, the different batch normalization layers used in the model for different dataset inputs can make it easier for the model to analyze feature differences between different datasets, improve the overall feature extraction capabilities of the model, and solve the problem of weak model feature extraction under small sample medical data. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 is a schematic diagram of a classification model for Alzheimer's disease according to an exemplary embodiment.

[0027] Figure 2 It is a schematic diagram of reconstructing the residual block BB-CBAM of the Resnet network according to an exemplary embodiment. DETAILED DESCRIPTION

[0028] The present invention will be described in detail below based on the accompanying drawings and preferred embodiments, and the purpose and effects of the present invention will become more clear. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0029] The present invention is based on a classification method for Alzheimer's disease enhanced by a collaborative learning method. The classification model of Alzheimer's disease constructed by this method is modified and optimized based on Resnet18 as the backbone network to adapt to the special scenario of medical data sets with unbalanced data distribution under small samples. The Resnet network adopts a residual network structure, which can solve the problem of gradient diffusion that is prone to occur in deep networks during back propagation. Taking into account the particularity of medical imaging, unbalanced data distribution and small amount of data, the focus is on improving the generalization ability of the model under AD data sets. The overall framework of the model is as follows Figure 1 At the same time, in order to improve the generalization ability of the model, a residual connected CBAM attention module is added at the end of the residual block (BasicBlock, BB) in the Resnet18 network. The use of the attention module can improve the model's ability to extract features from multiple angles. The modified and optimized model is named BB-CBAM layer, and the specific structure is as follows Figure 2 shown.

[0030] The Alzheimer's disease classification method enhanced by the collaborative learning method of the present invention specifically comprises the following steps:

[0031] Step 1: Construct a classification model for Alzheimer's disease, which uses Resnet18 as the backbone network and modifies the single fully connected layer of Resnet18 into at least two fully connected layers, which are used for Alzheimer's disease and natural image classification respectively; at the same time, the batch normalization layer at all positions of Resnet18 is modified from one to at least two, and finally an attention mechanism module is added to the residual block structure in Resnet18; the number of the fully connected layers is equal to the number of normalization layers at each position;

[0032] As one implementation method, there are two fully connected layers, the normalization layer at each position is changed from one to two, and the natural image dataset uses the cat and dog dataset. Figure 1 As shown, the model will remove the last layer of 1000-class fully connected layer classifiers of the original Resnet18, directly obtain 512 feature information, and add two fully connected layer classifiers corresponding to Alzheimer's disease and cat and dog classification in the layer after the feature output. When the feature is output, it is judged according to the input data set, and different fully connected layer classifiers are selected to generate different model predictions. At the same time, due to the different data distributions between different data sets, in order to better improve the model's ability to extract features, all BN layers of the original Resnet network are modified, and multiple BN layers are created in the position of the original BN layer, that is, the same number of BN layers as the data set are created, and different BN layers are selected according to different input data sets during model training and prediction. Finally, the CBAM attention mechanism module is appended to the residual block structure in the Resnet network, as shown in Figure 2 shown.

[0033] This embodiment only uses two data sets as examples. In theory, multiple data sets can be used for collaborative enhancement. The reason is that the model of the present invention uses multiple independent BN layers to normalize each data set, ensuring the characteristics between each data set. On the other hand, multiple data sets pass through the same network model (convolutional layer), etc., ensuring that convolution can learn the commonalities of multiple data sets. In theory, multiple data sets can be used for collaborative enhancement.

[0034] Among them, the specific formula for selecting different BN layers according to different input data sets is as follows:

[0035]

[0036]

[0037] Among them, BN i Represents the i-th normalization layer at a certain position of Resnet18, represents the jth input of the ith data set, μ and ε are the mean and variance of the input tensor calculated along the channel direction, respectively. ε in the formula is a constant to make the denominator non-zero; After standardization According to D i Different, choose different BN i , that is, different γ i , β i The added variable γ and translation variable β are trainable parameters of the BN layer, and their main function is to enhance the normalized expression ability.

[0038] Step 2: Select Alzheimer's disease image data and natural images, normalize them, and construct training sets and test sets for Alzheimer's disease and natural images, respectively. The specific steps include the following:

[0039] (1) For example, the Alzheimer's disease imaging data is selected from the Alzheimer's Disease Neuroimaging Project dataset, and the natural image is selected from the cat and dog dataset in the Kaggle competition network. A total of 431 T1-weighted structural MRI scans were collected from the ADNI database, from 200 AD patients and 231 normal subjects. The magnetic resonance image dataset of a given subject includes the magnetic resonance image and the subject's label, and the magnetic resonance image dataset is denoted as D = {(d i ,c j )|i∈[1,R],j∈L}, where d represents the magnetic resonance image, c represents the corresponding label of the subject, R represents the number of samples, i represents the sample subscript, j represents the label subscript, and L=[1,2] represents the label space, which represents the three categories of NC and AD respectively;

[0040] (2) The acquired magnetic resonance image set D is first normalized into a standard template space, and then the data is segmented into gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF);

[0041] (3) Based on the data obtained by segmentation in step (2), gray matter data with rich brain information is selected, and then smoothed (FMWH = 4 mm) to obtain a higher signal-to-noise ratio;

[0042] (4) Define the gray matter image obtained in step (3) as I, and divide the preprocessed data set into training set D according to the ratio of 8:2. train ={(d i ,c j )|i∈[1,M],j∈[1,2]} and the test set D test ={(d i ,c j)|i∈[1,N],j∈[1,2]}, where M is the total number of training set data and N is the total number of test set data.

[0043] Step 3: Alternately input the Alzheimer's disease training set and the natural image training set constructed in step 2 into the Alzheimer's disease classification model constructed in step 1, select the corresponding normalization layer and fully connected layer according to the different input training sets, and collaboratively train the model parameters. During the training process, sum up the losses of all training sets and perform back propagation. The predicted probabilities of multiple brain image slices are averaged by the slice voting method, and the average is used as the final output of the model to retain the optimal model for the classification of Alzheimer's disease;

[0044] The slice voting method is to pass all n slices of a subject through the same model to obtain a set of predicted values ​​(y1, y2, ..., y n ), and then calculate the mean μ of a set of prediction results, and the mean is used as the prediction result result of this subject.

[0045]

[0046] The use of a slice voting method that simulates doctors' film reading can greatly help improve the prediction stability and accuracy of the model.

[0047] As one of the implementation methods, during the model training process, the SGD optimizer is used to optimize the model training process. During the model training, different data sets D are alternately i Select a data set to construct a batch of samples and send them into the model to obtain the forward propagation loss. When the loss of each data set through the entire model is obtained, the losses are summed and back-propagated. The training uses the improved classification cross entropy loss function based on multiple data sets to calculate the training loss Loss. If the Loss is smaller, it means that the feature discrimination extracted by the trained model is greater, and the classification performance of the model on the training data is better.

[0048]

[0049] Where n is the data set D i , classes represents the number of categories, y represents the category variable (0 or 1), if the category is the same as the sample category, it is 1, otherwise it is 0. p represents the predicted probability that the sample belongs to category k after softmax calculation. When saving the model, select the model with the smallest Alzheimer's disease loss to save.

[0050] In order to test the various indicators of the model, the Alzheimer's disease imaging data and the test set data of cat and dog data were input into the optimal model for Alzheimer's disease classification to obtain the corresponding classification results. The model was evaluated by the following accuracy (ACC), precision (PRE), sensitivity (SEN), F1-Score and ROC curve, where the ROC curve was compared using its area under the line (AUC).

[0051]

[0052]

[0053]

[0054]

[0055] Among them, TP, TN, FP, and FN respectively represent the number of true positives, true negatives, false positives, and false negatives of the model for a given set of data classification. Accuracy is also called precision, which indicates the probability that the model predicts correctly; precision is also called positive predictive value (PPV), which is the ratio of the number of predicted positive cases to the number of actual positive cases; sensitivity is also called recall, which is the probability of being correctly predicted among the predicted positive cases, reflecting the sensitivity of the model to different categories of features; the larger the value of the three indicators, the better the model performance.

[0056] In this embodiment, the server graphics card device uses NVIDIA Tesla V100, the deep learning framework uses Pytorch, and the network training uses a small batch stochastic gradient descent algorithm, where the learning rate is set to 0.01, the momentum is 0.9, the weight decay is 0.0001, and the batch size is 16. The total number of iterations (epochs) is 100, and the learning rate continues to decay with the increase of the number of iterations epoch. epoch+1 =lr epoch ×0.9.

[0057] The best model saved from training, when applied to the test set, showed a best binary classification accuracy of 88.3%. The detailed classification performance parameters of the model on the test set are shown in Table 1. As can be seen from Table 1, the overall performance has achieved satisfactory results, and the sensitivity to the disease is high, which makes the auxiliary diagnosis more reliable. This is because the model trained through cross-dataset collaboration can effectively analyze the feature differences between different data sets, thereby efficiently extracting important feature information from the image and better capturing the structural information changes in medical images, which is of great significance in medical image analysis. From the comprehensive evaluation standard F1-Score, it can be seen that the model has excellent performance in binary classification tasks.

[0058] Table 1 Model accuracy

[0059] Model ACC REC PRE F1 Control group 0.802 0.717 0.891 0.795 Resnet34 0.795 0.825 0.874 0.849 Alexnet 0.813 0.804 0.840 0.822 Vgg11 0.848 0.891 0.836 0.863 Vgg16 0.845 0.877 0.822 0.849 The present invention proposes 0.883 0.956 0.846 0.897

[0060] As can be seen from Table 1, the method proposed in the present invention has been significantly improved compared with the control group and other network model methods, especially with a 6% increase in recall rate, which is a good performance, and a 4% increase in accuracy, which is a good overall performance. The experimental results verify that under the same environment, the collaborative learning method can improve the generalization ability of model training, and the attention mechanism combined with the convolutional neural network has a certain degree of promotion effect on the final classification prediction.

[0061] Those skilled in the art can understand that the above are only preferred examples of the invention and are not intended to limit the invention. Although the invention is described in detail with reference to the above examples, those skilled in the art can still modify the technical solutions recorded in the above examples or replace some of the technical features therein with equivalents. Any modification, equivalent replacement, etc. made within the spirit and principle of the invention shall be included in the protection scope of the invention.

Claims

1. A classification method for Alzheimer's disease enhanced by collaborative learning method, characterized in that: The steps include: Step 1: Construct a classification model for Alzheimer's disease, which uses Resnet18 as the backbone network and modifies the single fully connected layer of Resnet18 into at least two fully connected layers, which are used for Alzheimer's disease and natural image classification respectively; at the same time, the batch normalization layer at all positions of Resnet18 is modified from one to at least two, and finally an attention mechanism module is added to the residual block structure in Resnet18; the number of the fully connected layers is equal to the number of normalization layers at each position; Step 2: Select Alzheimer's disease image data and natural images, and normalize them to construct a training set for Alzheimer's disease and a training set for natural images respectively; Step 3: Alternately input the Alzheimer's disease training set and the natural image training set constructed in step 2 into the Alzheimer's disease classification model constructed in step 1, select the corresponding normalization layer and fully connected layer according to the different input training sets, and collaboratively train the model parameters. During the training process, sum up the losses of all training sets and perform back propagation. The predicted probabilities of multiple brain image slices are averaged by the slice voting method, and the average is used as the final output of the model to retain the optimal model for the classification of Alzheimer's disease; Step 4: The Alzheimer's disease image data to be classified is also normalized in the same way as in step 2, and then input into the optimal model for Alzheimer's disease classification, and the model outputs the corresponding classification results.

2. The classification method for Alzheimer's disease based on collaborative learning method enhancement according to claim 1, characterized in that: In step 3, the final output features of the model will select different fully connected layers according to the different input data sets to obtain the final classification results: y i (x i )=classify i (x i ) Among them, x i is the data of the i-th data set, classify i To select different classifiers according to different data sets, different classification results y are finally obtained. i .

3. The classification method for Alzheimer's disease based on collaborative learning method enhancement according to claim 1, characterized in that: In the step 3, according to the input training set D i The normalization layers corresponding to different choices are specifically expressed as follows: Among them, BN i Represents the i-th normalization layer at a certain position of Resnet18, represents the jth input of the i-th data set, μ and ε are the mean and variance of the input tensor calculated along the channel direction, and ε in the formula is a constant; After standardization According to D i Different, choose different BN i , that is, different γ i , β i .

4. The classification method for Alzheimer's disease based on collaborative learning method enhancement according to claim 1, characterized in that: In step 4, when training the model, different training sets D are used alternately. i Select a dataset to construct a batch of samples and feed them into the model to obtain the forward propagation loss. When the loss of each dataset through the entire classification model is obtained, the loss is summed and back-propagated. Where n is the data set D i , classes represents the number of categories, y represents the category variable, y is 0 or 1, if the category is the same as the sample category, y is 1, otherwise y is 0; p represents the predicted probability that the sample belongs to category k after softmax calculation; the model with the smallest Alzheimer's disease loss is selected as the optimal model for the classification of Alzheimer's disease.

5. The classification method for Alzheimer's disease based on collaborative learning method enhancement according to claim 1, characterized in that: In step 3, during the back propagation process, the SGD optimizer is used to optimize the model parameters.

Images