Alzheimer disease classification method based on MRI image

Abstract

The invention discloses an MRI (Magnetic Resonance Imaging) image-based Alzheimer's disease classification method, which comprises the following steps of: selecting an sMRI image of a subject diagnosed for the first time, and preprocessing to obtain an sMRI grey matter image; segmenting the sMRI grey matter image into a plurality of regional images, and encoding the regional images corresponding to the half brain through a regional feature extraction network to obtain feature vector groups of the left half brain and the right half brain; inputting the feature vector groups of the left and right half brains into Transform modules of the left and right half brains, splicing the obtained features of the left and right half brains, and inputting the spliced features of the left and right half brains into a Transform module of the whole brain to obtain whole brain features; calculating cosine loss and hybrid cross entropy loss to obtain hybrid loss, and optimizing model network parameters through the hybrid loss; and based on the optimized model network, obtaining a left and right half brain branch prediction result and a whole brain prediction result, and completing Alzheimer's disease classification. According to the method, the reliability of classification and auxiliary diagnosis is improved, and the AD classification accuracy is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing, and in particular relates to an Alzheimer's disease classification method based on MRI images. Background Art

[0002] Alzheimer's disease (AD) is a degenerative disorder of the central nervous system characterized by memory loss, language dysfunction, and cognitive and behavioral impairments. It accounts for approximately 60%-70% of dementia cases worldwide. Its progressive neurodegenerative process not only leads to the loss of independence but also significant emotional and social impairment, placing a heavy burden on families and public healthcare systems. According to the World Health Organization, with the accelerated aging of the global population, the number of AD patients is increasing at a rate of approximately 10 million new cases annually, and the total number of patients is expected to exceed 150 million by 2050. In this context, early and accurate diagnosis is crucial for delaying disease progression and optimizing targeted intervention strategies. Advances in imaging technology have led to the emergence of various neuroimaging techniques, including structural magnetic resonance imaging (sMRI). As a non-invasive, radiation-free structural imaging method, sMRI is currently the most widely used imaging tool in clinical practice. High-resolution three-dimensional imaging reveals typical morphological features such as hippocampal volume atrophy, decreased temporal lobe gray matter density, and enlarged lateral ventricles. sMRI allows physicians to visualize these distinct structural changes in the brain to assess AD. The two imaging techniques mentioned above provide more evidence for the screening and early diagnosis of AD. Therefore, how to effectively utilize these imaging features to improve AD diagnosis performance using current deep learning technology has become a research hotspot in recent years.

[0003] MRI-based cognitive score estimation has important clinical value in dementia, helping to assess the pathological stage of AD and predict disease progression. Existing algorithms generally lack the ability to model long-range dependencies across the entire brain, making it difficult to effectively integrate local morphological features with global functional associations, resulting in limited ability to express features of diffuse brain atrophy. Summary of the Invention

[0004] In response to the above-mentioned deficiencies in the existing technology, the present invention provides an Alzheimer's disease classification method based on MRI images, which solves the problem that the existing technology lacks the ability to model long-range dependencies within the entire brain and is difficult to effectively integrate local morphological features with global functional associations.

[0005] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: a method for classifying Alzheimer's disease based on MRI images, comprising the following steps: S1. Based on the public dataset ADNI, the sMRI images of the subjects at their first diagnosis were preprocessed to obtain sMRI gray matter images. S2. Segment the sMRI gray matter image into several regional images, encode the regional images of the corresponding hemispheres through the regional feature extraction network of the left and right hemispheres, and concatenate the vectors of the left and right hemisphere encodings to obtain the feature vector group of the left and right hemispheres; S3: Input the feature vector groups of the left and right hemispheres into the Transformer modules of the left and right hemispheres to obtain more representative features of the left and right hemispheres. The left and right hemisphere features are then concatenated and input into the Transformer module of the whole brain to obtain the whole brain features. S4. Calculate the cosine loss based on the feature vector groups of the left and right hemispheres, calculate the mixed cross entropy loss based on the left and right hemisphere features and the whole-brain features, add the cosine loss and the mixed cross entropy loss according to the set ratio to obtain the mixed loss, and optimize the model network parameters through the mixed loss. The model network includes the regional feature extraction network of the left and right hemispheres, the Transformer module of the left and right hemispheres, and the Transformer module of the whole brain; S5. Based on the optimized model network, new sMRI images are obtained, and S1 to S4 are executed to generate new left and right brain features and new whole-brain features. The left and right brain branch prediction results are obtained based on the new left and right brain features, and the whole-brain prediction results are obtained based on the new whole-brain features to complete the Alzheimer's disease classification.

[0006] Furthermore: In S1, the method for preprocessing the sMRI image is specifically as follows: The sMRI images were subjected to bias correction, template registration and normalization processing, and gray matter tissue was extracted from the sMRI images. The resolution of the generated sMRI gray matter images was 105×125×105.

[0007] Furthermore: in S2, the regional feature extraction network structures of the left and right hemispheres are the same, both including a first three-dimensional convolutional layer, a second three-dimensional convolutional layer, a maximum pooling layer, a third three-dimensional convolutional layer, a fourth three-dimensional convolutional layer and a global average pooling layer connected in sequence; Among them, the size of the first 3D convolutional layer is 4×4×4, and the output channels are 32. The size of the second 3D convolutional layer is 3×3×3, and the output channels are 64. The size of the maximum pooling layer is 2×2×2. The size of the third and fourth 3D convolutional layers is 3×3×3, and the output channels are 128. The size of the global average pooling layer is 2×2×2.

[0008] Further: S2 includes the following sub-steps: S21. Segment the sMRI gray matter image into several non-overlapping regional images with a resolution of 25×25×25; S22, assigning regional feature extraction networks with unshared parameters to the regional images of the left and right hemispheres, inputting the regional images of the left and right hemispheres into the corresponding regional feature extraction networks, respectively, encoding the regional images of the corresponding hemispheres, and encoding the encoding vectors with a dimension of 128; S23. Concatenate the vectors encoded by the left and right hemispheres to obtain a set of feature vectors of the left and right hemispheres.

[0009] Furthermore: In S3, the Transformer module structure of the left hemisphere, right hemisphere, and whole brain is the same, including: input layer, first normalization layer, multi-head self-attention module, second normalization layer, and multi-layer perceptron layer; Among them, the input layer is connected to the first normalization layer and the multi-head self-attention module in sequence. The output of the multi-head self-attention module is spliced ​​with the input of the first normalization layer through a jump connection. The splicing result is input into the second normalization layer and the multi-layer perceptron layer in sequence. The output of the multi-layer perceptron layer is spliced ​​with the input of the second normalization layer through a jump connection, and the splicing result is used as the output of the Transformer module.

[0010] The beneficial effect of the above further scheme is: the present invention designs a method for enhancing the representation ability of hemi-brain features based on the Transformer module, and through the self-attention mechanism sub-module of the Transformer module, it fuses the feature information of other regions for each regional feature, enhances the representation ability of each feature vector, and outputs a hemi-brain feature with the same scale as the input but with better representation ability.

[0011] Further: S3 includes the following sub-steps: S31, assigning left and right hemisphere feature vector groups to Transformer modules of the left and right hemispheres, and inputting the left and right hemisphere feature vector groups into the corresponding hemisphere Transformer modules, respectively, to obtain more representative left and right hemisphere features; S32, the left and right hemisphere features are concatenated and input into the whole-brain Transformer module, and the whole-brain features are connected through the whole-brain Transformer module to obtain the whole-brain features; Among them, the Transformer modules of the left and right hemispheres and the Transformer modules of the whole brain adopt a training method with non-shared parameters.

[0012] The beneficial effect of the above further scheme is that the present invention adopts multiple Transformer modules to better learn the characteristics of each modality image through non-shared parameter training, further improving the auxiliary diagnosis performance of Alzheimer's disease based on multimodal image feature fusion.

[0013] Further: In S4, calculate the cosine loss The specific expression is: Where, For the i The feature vector of the right hemisphere, For the i The eigenvectors of the left hemisphere, N is the total number of hemibrain eigenvectors; Calculate hybrid cross entropy loss The specific expression is: Where, and They are the regional feature extraction network and Transformer module for the left and right hemispheres respectively. A Transformer module for the whole brain; Mixing loss The specific expression is: .

[0014] The beneficial effect of the above further scheme is: the present invention designs a hybrid loss based on the hybrid cross entropy loss of cosine similarity and multi-branch prediction, which alleviates the gradient vanishing problem of long-distance networks and further improves network performance and convergence speed.

[0015] Furthermore: In S5, the method for classifying Alzheimer's disease is specifically as follows: The new left and right hemisphere features are input into the linear layer to obtain the left and right hemisphere branch prediction results. The new whole-brain features are input into the fully connected layer and the softmax layer in sequence to obtain the whole-brain prediction results. Alzheimer's disease classification is completed based on the obtained left and right hemisphere branch prediction results and the whole-brain prediction results.

[0016] The beneficial effects of the present invention are: (1) The present invention provides an Alzheimer's disease classification method based on MRI images. It proposes a regional feature extraction network for multi-scale 3D local feature encoding, brain hemispheric symmetry analysis, and a model network enhanced by the Transformer module. Through the multi-level feature fusion and adaptive optimization mechanism of left and right hemisphere features, it breaks through the technical bottleneck of the existing technology's lack of modeling ability for long-range dependencies within the whole brain, difficulty in effectively integrating local morphological features with global functional associations, and limited ability to express diffuse brain atrophy characteristics, thereby improving the accuracy and clinical applicability of early AD diagnosis. Evaluation on the ADNI baseline dataset shows that the method of the present invention performs well in estimating cognitive scores.

[0017] (2) The present invention realizes AD auxiliary diagnosis based on multiple sMRI gray matter images. First, considering the characteristics of brain images, in order to reduce the loss of detail information in the whole-brain convolution feature extraction, the sMRI image is divided into multiple fixed-size, closely arranged and non-overlapping regional images before feature extraction, and a regional feature extraction network with the same structure but non-shared parameters is used for the left and right brains to perform feature encoding for each sub-region. Secondly, in order to strengthen the representation ability of the feature vector group and not miss the key information of the whole brain, the feature expression of the left and right brains is strengthened by an independent Transformer module, and the Transformer module is used again to learn the connection between the features of different regions of the left and right hemispheres to obtain the whole-brain feature expression. Finally, the final classification prediction is obtained by expressing the left and right hemisphere features and the whole-brain features, which improves the reliability of classification and auxiliary diagnosis, and thus improves the classification accuracy of AD. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a flow chart of an Alzheimer's disease classification method based on MRI images of the present invention.

[0019] Figure 2 This is a principle block diagram of the model network structure of the present invention.

[0020] Figure 3 This is the regional feature extraction network structure diagram of the present invention.

[0021] Figure 4 This is the Transformer module structure diagram of the present invention. DETAILED DESCRIPTION

[0022] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0023] like Figure 1 As shown, in one embodiment of the present invention, a method for classifying Alzheimer's disease based on MRI images includes the following steps: S1. Based on the public dataset ADNI, the sMRI images of the subjects at their first diagnosis were preprocessed to obtain sMRI gray matter images. S2. Segment the sMRI gray matter image into several regional images, encode the regional images of the corresponding hemispheres through the regional feature extraction network of the left and right hemispheres, and concatenate the vectors of the left and right hemisphere encodings to obtain the feature vector group of the left and right hemispheres; S3: Input the feature vector groups of the left and right hemispheres into the Transformer modules of the left and right hemispheres to obtain more representative features of the left and right hemispheres. The left and right hemisphere features are then concatenated and input into the Transformer module of the whole brain to obtain the whole brain features. S4. Calculate the cosine loss based on the feature vector groups of the left and right hemispheres, calculate the mixed cross entropy loss based on the left and right hemisphere features and the whole-brain features, add the cosine loss and the mixed cross entropy loss according to the set ratio to obtain the mixed loss, and optimize the model network parameters through the mixed loss. The model network includes the regional feature extraction network of the left and right hemispheres, the Transformer module of the left and right hemispheres, and the Transformer module of the whole brain; S5. Based on the optimized model network, new sMRI images are obtained, and S1 to S4 are executed to generate new left and right brain features and new whole-brain features. The left and right brain branch prediction results are obtained based on the new left and right brain features, and the whole-brain prediction results are obtained based on the new whole-brain features to complete the Alzheimer's disease classification.

[0024] In this embodiment, the overall model network of the present invention is as follows Figure 2 As shown, the present invention designs a Transformer module and a regional feature extraction network in the model network to achieve multi-level feature fusion of the left and right hemispheres.

[0025] In S1, the public dataset ADNI was used to formulate a non-leaked partitioned dataset. A total of sMRI images of 447 subjects in the ADNI database were used, including 234 NC subjects and 213 AD subjects. This embodiment divides the dataset by subject number at a ratio of 70%, 15%, and 15% based on the number of subjects. The first n-1 numbered subjects are used for network training, half of the data after n is used for verification, and the other half is used for testing. All subjects are not from the same person, which avoids inaccurate test results due to data leakage. The two modal images were preprocessed using the Clinica software platform. After bias correction, template alignment, and pixel normalization, the preprocessed images were obtained. The resolution of the preprocessed images was 105×125×105.

[0026] In S1, the method for preprocessing sMRI images is as follows: The sMRI images were subjected to bias correction, template registration and normalization processing, and gray matter tissue was extracted from the sMRI images. The resolution of the generated sMRI gray matter images was 105×125×105.

[0027] like Figure 3As shown in Figure 2, in S2, the regional feature extraction network structures of the left and right hemispheres are the same, both including the first 3D convolution layer, the second 3D convolution layer, the maximum pooling layer, the third 3D convolution layer, the fourth 3D convolution layer and the global average pooling layer connected in sequence; Among them, the size of the first 3D convolutional layer is 4×4×4, and the output channels are 32. The size of the second 3D convolutional layer is 3×3×3, and the output channels are 64. The size of the maximum pooling layer is 2×2×2. The size of the third and fourth 3D convolutional layers is 3×3×3, and the output channels are 128. The size of the global average pooling layer is 2×2×2.

[0028] In this embodiment, a shallow 3D convolutional neural network with the same structure and no shared parameters is assigned to the left and right hemispheres as the regional feature extraction networks for the left and right hemispheres, and the regional images are input into the regional feature extraction network. For each regional image with a size of 25×25×25, convolution is first performed through a first three-dimensional convolutional layer with a size of 4×4×4, and then convolution is performed through a second three-dimensional convolutional layer with a size of 3×3×3. The output result is subjected to a 2×2×2 maximum pooling layer and input into the third and fourth three-dimensional convolutional layers again. The obtained feature map is globally average pooled to obtain a 128-dimensional vector. All the vectors obtained for each hemisphere are concatenated to obtain a feature vector group representing the features of the hemisphere, and the average cosine similarity of the two groups of feature vectors is calculated as the loss.

[0029] S2 includes the following sub-steps: S21. Segment the sMRI gray matter image into several fixed-size, closely arranged, and non-overlapping regional images with a resolution of 25×25×25. S22, assigning regional feature extraction networks with unshared parameters to the regional images of the left and right hemispheres, inputting the regional images of the left and right hemispheres into the corresponding regional feature extraction networks, respectively, encoding the regional images of the corresponding hemispheres, and encoding the encoding vectors with a dimension of 128; S23. Concatenate the vectors encoded by the left and right hemispheres to obtain a set of feature vectors of the left and right hemispheres.

[0030] In this embodiment, after each regional image is input into the regional feature extraction network, a 128-dimensional vector is obtained. All vectors obtained from each hemisphere are concatenated to obtain a feature vector group representing the left and right hemispheres of the brain.

[0031] like Figure 4 As shown in Figure 3, the Transformer module structures of the left hemisphere, right hemisphere, and whole brain are the same, including: input layer, first normalization layer, multi-head self-attention module, second normalization layer, and multi-layer perceptron layer; Among them, the input layer is connected to the first normalization layer and the multi-head self-attention module in sequence. The output of the multi-head self-attention module is spliced ​​with the input of the first normalization layer through a jump connection. The splicing result is input into the second normalization layer and the multi-layer perceptron layer in sequence. The output of the multi-layer perceptron layer is spliced ​​with the input of the second normalization layer through a jump connection, and the splicing result is used as the output of the Transformer module.

[0032] In this embodiment, the workflow of the Transformer module is as follows: The input vector of the Transformer module first enters the first normalization layer, which normalizes the features and improves the stability of model training. After normalization, it enters the multi-head self-attention module. Through multiple independent attention parallel calculations, it captures the dependencies between different positions in the sequence. The output of the multi-head self-attention module is concatenated with the input of the first normalization layer and passed to the second normalization layer to further adjust the feature distribution. After processing in the second normalization layer, the features enter the multi-layer perceptron (MLP) layer, ultimately resulting in the output of the Transformer module.

[0033] S3 includes the following sub-steps: S31, assigning left and right hemisphere feature vector groups to Transformer modules of the left and right hemispheres, inputting the left and right hemisphere feature vector groups into the corresponding hemisphere Transformer modules, and outputting hemisphere feature vector groups of the same scale but with better representation ability, which are used as the left and right hemisphere features with better representation ability; S32, the left and right hemisphere features are concatenated and input into the whole-brain Transformer module, and the whole-brain features are connected through the whole-brain Transformer module to obtain the whole-brain features; Among them, the Transformer modules of the left and right hemispheres and the Transformer modules of the whole brain adopt a training method with non-shared parameters.

[0034] In this embodiment, the present invention S3 designs a method for enhancing the characterization capability of hemibrain features based on the Transformer module. Based on the obtained multiple regional feature vectors extracted from the hemibrain, a group of n×128 left and right hemibrain feature vectors is obtained by merging the feature vectors of the left and right hemibrains. The self-attention mechanism submodule of the Transformer module fuses the feature information of other regions for each regional feature, strengthening the characterization capability of each feature vector, and outputting a hemibrain feature with the same input scale but greater characterization capability. Because the Transformer module re-expresses a group of vectors after fusion, each feature vector contains more feature information of its related vectors, which is conducive to further improving the network classification performance. After merging the two hemibrain-enhanced feature vector groups, the whole-brain Transformer module is used to construct the connectivity of the whole-brain features, generating a whole-brain feature that more effectively expresses the whole-brain information. The vector group of the whole-brain feature associates the features of different sub-regions of the left and right brain, and combines multiple regional features to achieve a more effective expression of the whole-brain feature. Therefore, this embodiment uses multiple Transformer modules through a non-shared parameter training method to better learn the features of each modality image, further improving the performance of Alzheimer's disease auxiliary diagnosis based on multimodal image feature fusion.

[0035] In S4, calculate the cosine loss The specific expression is: Where, For the i The feature vector of the right hemisphere, For the i The eigenvectors of the left hemisphere, N is the total number of hemibrain eigenvectors; Calculate hybrid cross entropy loss The specific expression is: Where, and They are the regional feature extraction network and Transformer module for the left and right hemispheres respectively. A Transformer module for the whole brain; Mixing loss The specific expression is: .

[0036] In this embodiment, the present invention designs a hybrid loss based on the hybrid cross entropy loss of cosine similarity and multi-branch prediction to alleviate the gradient vanishing problem of long-distance networks and further improve network performance and convergence speed.

[0037] In S5, the method for classifying Alzheimer's disease is as follows: The new left and right hemisphere features are input into the linear layer to obtain the left and right hemisphere branch prediction results. The new whole-brain features are input into the fully connected layer and the softmax layer in sequence to obtain the whole-brain prediction results. Alzheimer's disease classification is completed based on the obtained left and right hemisphere branch prediction results and the whole-brain prediction results.

[0038] In this embodiment, in order to evaluate the classification performance of the overall model network of the present invention for AD patients and cognitively normal subjects (NC), three evaluation indicators are introduced according to actual applications: accuracy Accuracy , sensitivity Sensitivity and specificity Specificity , as shown below: Where, TP is the true positive, representing the number of AD samples predicted to be AD, FP is a false positive, representing the number of NC samples predicted to be AD, FN is a false negative, representing the number of AD samples predicted to be NC, TN is the true negative, representing the number of NC samples predicted to be NC.

[0039] Accuracy refers to the percentage of correctly classified subjects out of the total number of subjects in the test set, and is used to evaluate the model's overall discriminative ability in classifying images of AD and NC subjects; sensitivity is a measure of missed diagnosis; and specificity is an important indicator for diagnosing a disease. Table 1 shows the comparison results between the network of the present invention and a network based on whole-brain 3D convolution plus Transformer feature enhancement (3DCNN+VIT) and a network based on regional segmentation for 3D feature extraction (3D Patch CNN), reflecting that the method of the present invention has obvious advantages in the diagnostic performance of Alzheimer's disease.

[0040] Table 1 Comparison of diagnostic performance between the method of the present invention and other network methods In the description of the present invention, it should be understood that the terms "center", "thickness", "upper", "lower", "horizontal", "top", "bottom", "inner", "outer", "radial", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and "third" are used for descriptive purposes only and cannot be understood as indicating or implying the relative importance or the number of technical features implicitly specified. Therefore, the features defined by "first", "second", and "third" may explicitly or implicitly include one or more of such features.

Claims

1. A method for classifying Alzheimer's disease based on MRI images, characterized in that: The following steps are involved: S1. Based on the public dataset ADNI, the sMRI images of the subjects at their first diagnosis were preprocessed to obtain sMRI gray matter images. S2. Segment the sMRI gray matter image into several regional images, encode the regional images of the corresponding hemispheres through the regional feature extraction network of the left and right hemispheres, and concatenate the vectors of the left and right hemisphere encodings to obtain the feature vector group of the left and right hemispheres; S3: Input the feature vector groups of the left and right hemispheres into the Transformer modules of the left and right hemispheres to obtain more representative features of the left and right hemispheres. The left and right hemisphere features are then concatenated and input into the Transformer module of the whole brain to obtain the whole brain features. S4. Calculate the cosine loss based on the feature vector groups of the left and right hemispheres, calculate the mixed cross entropy loss based on the left and right hemisphere features and the whole-brain features, add the cosine loss and the mixed cross entropy loss according to the set ratio to obtain the mixed loss, and optimize the model network parameters through the mixed loss. The model network includes the regional feature extraction network of the left and right hemispheres, the Transformer module of the left and right hemispheres, and the Transformer module of the whole brain; S5. Based on the optimized model network, new sMRI images are obtained, and S1 to S4 are executed to generate new left and right brain features and new whole-brain features. The left and right brain branch prediction results are obtained based on the new left and right brain features, and the whole-brain prediction results are obtained based on the new whole-brain features to complete the Alzheimer's disease classification.

2. The Alzheimer's disease classification method based on MRI images according to claim 1, characterized in that: In S1, the method for preprocessing sMRI images is as follows: The sMRI images were subjected to bias correction, template registration and normalization processing, and gray matter tissue was extracted from the sMRI images. The resolution of the generated sMRI gray matter images was 105×125×105.

3. The Alzheimer's disease classification method based on MRI images according to claim 1, characterized in that: In S2, the regional feature extraction network structure of the left and right hemispheres is the same, both including the first 3D convolution layer, the second 3D convolution layer, the maximum pooling layer, the third 3D convolution layer, the fourth 3D convolution layer and the global average pooling layer connected in sequence; Among them, the size of the first 3D convolutional layer is 4×4×4, and the output channels are 32. The size of the second 3D convolutional layer is 3×3×3, and the output channels are 64. The size of the maximum pooling layer is 2×2×2. The size of the third and fourth 3D convolutional layers is 3×3×3, and the output channels are 128. The size of the global average pooling layer is 2×2×2.

4. The Alzheimer's disease classification method based on MRI images according to claim 3, characterized in that: S2 includes the following sub-steps: S21. Segment the sMRI gray matter image into several non-overlapping regional images with a resolution of 25×25×25; S22, assigning regional feature extraction networks with unshared parameters to the regional images of the left and right hemispheres, inputting the regional images of the left and right hemispheres into the corresponding regional feature extraction networks, respectively, encoding the regional images of the corresponding hemispheres, and encoding the encoding vectors with a dimension of 128; S23. Concatenate the vectors encoded by the left and right hemispheres to obtain a set of feature vectors of the left and right hemispheres.

5. The Alzheimer's disease classification method based on MRI images according to claim 1, characterized in that: In S3, the Transformer module structure of the left hemisphere, right hemisphere, and whole brain is the same, including: input layer, first normalization layer, multi-head self-attention module, second normalization layer, and multi-layer perceptron layer; Among them, the input layer is connected to the first normalization layer and the multi-head self-attention module in sequence. The output of the multi-head self-attention module is spliced ​​with the input of the first normalization layer through a jump connection. The splicing result is input into the second normalization layer and the multi-layer perceptron layer in sequence. The output of the multi-layer perceptron layer is spliced ​​with the input of the second normalization layer through a jump connection, and the splicing result is used as the output of the Transformer module.

6. The Alzheimer's disease classification method based on MRI images according to claim 5, characterized in that: S3 includes the following sub-steps: S31, assigning left and right hemisphere feature vector groups to Transformer modules of the left and right hemispheres, and inputting the left and right hemisphere feature vector groups into the corresponding hemisphere Transformer modules, respectively, to obtain more representative left and right hemisphere features; S32, after splicing the left and right hemisphere features, input them into the whole-brain Transformer module, and construct the connectivity of the whole-brain features through the whole-brain Transformer module to obtain the whole-brain features; Among them, the Transformer modules of the left and right hemispheres and the Transformer modules of the whole brain adopt a non-shared parameter training method.

7. The Alzheimer's disease classification method based on MRI images according to claim 1, characterized in that: In S4, calculate the cosine loss The specific expression is: Where, For the i The feature vector of the right hemisphere, For the i The eigenvectors of the left hemisphere, N is the total number of hemibrain eigenvectors; Calculate hybrid cross entropy loss The specific expression is: Where, and They are the regional feature extraction network and Transformer module for the left and right hemispheres respectively. A Transformer module for the whole brain; Mixing loss The specific expression is: 。 8. The Alzheimer's disease classification method based on MRI images according to claim 1, characterized in that: In S5, the method for classifying Alzheimer's disease is as follows: The new left and right hemisphere features are input into the linear layer to obtain the left and right hemisphere branch prediction results. The new whole-brain features are input into the fully connected layer and the softmax layer in sequence to obtain the whole-brain prediction results. Alzheimer's disease classification is completed based on the obtained left and right hemisphere branch prediction results and the whole-brain prediction results.

Images