Alzheimer's disease process identification method and system based on domain adaptation

Through convolutional neural networks that combat learning and attention mechanisms, the problem of different data in different hospitals affecting model performance is solved, the accuracy and stability of Alzheimer's disease process recognition is improved, and important auxiliary indicators are provided for diagnosis.

CN119028586APending Publication Date: 2024-11-26HUAQIAO UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411155732.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-22
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

There are differences in Alzheimer's disease data collected from different hospitals or sites, which affects the identification performance of the model.

Method used

Using a domain adaptation-based method, by preprocessing and data augmenting brain image data, a convolutional neural network training model with adversarial learning and attention mechanisms can learn features with domain invariance and reduce the source domain. and the gap between the characteristics of the target domain.

Benefits of technology

It improves the accuracy of Alzheimer's disease process identification model, alleviates the impact of data differences between different sites on model performance, and serves as an auxiliary diagnostic indicator to help doctors formulate diagnostic plans as early as possible.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119028586A_ABST
    Figure CN119028586A_ABST
Patent Text Reader

Abstract

The invention discloses an Alzheimer's disease process identification method and system based on domain adaptation, and the method comprises the following steps: carrying out the preprocessing of brain image data, and obtaining a source domain image; performing data enhancement processing on the source domain image to obtain an enhanced source domain image as a target domain image; training an Alzheimer's disease process recognition model by using the target domain image and the source domain image; and utilizing the trained Alzheimer's disease process identification model to realize Alzheimer's disease process identification. The Alzheimer's disease process identification model comprises a feature encoder and a classifier, the feature encoder extracts features of an image to be identified, and the classifier performs classification and process identification based on the extracted features. According to the method, learning of a domain loss function constraint model and feature expression of aligning source domain and target domain features are provided, so that the model learns features with domain invariant properties, and the influence of data differences of different sites on the diagnosis performance of the model is relieved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence model technology, and in particular to a method and system for identifying the progression of Alzheimer's disease based on domain adaptation. Background Art

[0002] Currently, most structural MRI image data used in auxiliary diagnosis of Alzheimer's disease come from multiple sites. Data from multiple sites increase the sample size of the data and improve the performance of the model; however, these data are often acquired using different scanning protocols, scanners, etc., which leads to differences between data from different sites. In actual clinical diagnosis, the process of collecting patient data in different hospitals and the data collection process used for standards and training are difficult to be consistent. In the field of Alzheimer's disease recognition, the differences between data from different sites will significantly affect the recognition performance of the model. Summary of the invention

[0003] The purpose of the present invention is to solve the problem in the prior art that model prediction is affected by differences in data collected from different hospitals or different sites, and to improve the accuracy of model recognition.

[0004] The technical solution adopted by the present invention to solve the technical problem is: to provide a method for identifying the progression of Alzheimer's disease based on domain adaptation, comprising the following steps:

[0005] Preprocess the brain image data to obtain source domain images;

[0006] Perform data enhancement processing on the source domain image to obtain the enhanced source domain image as the target domain image;

[0007] Using target domain images and source domain images to train an Alzheimer's disease progression recognition model;

[0008] Using the trained Alzheimer's disease process identification model to achieve Alzheimer's disease process identification;

[0009] The Alzheimer's disease process recognition model includes a feature encoder and a classifier. The feature encoder uses a convolutional neural network combined with an attention mechanism to extract features of the image to be recognized, and the classifier classifies and recognizes the process based on the extracted features.

[0010] Preferably, the preprocessing of brain image data includes:

[0011] The brain imaging data were corrected for head motion, non-uniform intensity, Talairach space transformation, normalization, and skull stripping;

[0012] All data were realigned to MNI space using SPM12 software.

[0013] Preferably, the method of training the Alzheimer's disease progression recognition model using the target domain image and the source domain image includes:

[0014] Input the source domain image into the feature encoder to obtain the source domain image features;

[0015] Input the target domain image into the feature encoder to obtain the target domain image features;

[0016] Using source domain image features and target domain image features for adversarial learning, the feature encoder learns features with domain-invariant properties.

[0017] The weight information from the classifier is used to weight the source domain image features to enhance the expression of the source domain features, and the enhanced source domain features are aligned with the target domain features to improve the classifier's perception of pathological details.

[0018] Preferably, the feature encoder includes three improved residual blocks and a global average pooling layer, each improved residual block performs feature encoding on the input image to be identified, the outputs of the three improved residual blocks are input into the global average pooling layer together, and the output of the global average pooling layer is used as the extracted feature; each improved residual block includes a first 3×3×3 convolutional layer, a second 3×3×3 convolutional layer, a first coordinate attention layer, a third 3×3×3 convolutional layer, a fourth 3×3×3 convolutional layer and a second coordinate attention layer connected in sequence.

[0019] Preferably, the feature encoder extracts features of the image to be identified, and the image to be identified passes through a 7×7×7 convolution layer and a 3×3×3 maximum pooling layer in sequence before entering the feature encoder.

[0020] Preferably, the adversarial learning using source domain image features and target domain image features comprises the following steps:

[0021] A classifier is used to build a domain discriminator, and the source domain image features and the target domain image features are input into the domain discriminator as two categories;

[0022] A domain loss function is proposed, expressed as:

[0023]

[0024] Where i represents the data sequence number, N represents the number of data; y i D Represents data x i The field label, y i =1 indicates data x i From the source domain, y i =0 means data x ifrom the target domain; F represents the feature encoder, C D represents the domain discriminator;

[0025] By maximizing the domain loss function, the domain discriminator cannot determine whether the data comes from the source domain or the target domain, thereby narrowing the gap between the source domain features and the target domain features, and allowing the feature encoder to learn features with domain-invariant properties.

[0026] Preferably, the step of weighting the source domain image features by using the weight information from the classifier to strengthen the expression of the source domain features comprises the following steps:

[0027] Three pooling layers with filter sizes of (H, W, 1), (1, W, D) and (H, 1, D) are used to pool each channel of the source domain features to obtain feature maps that can perceive direction, where H, W and D are the height, width and depth of the source domain features respectively;

[0028] The obtained feature maps are concatenated in the same spatial dimension, input into a 1×1×1 convolution layer, BN normalization and ReLU activation function in sequence, and then split into three vectors. After passing through a 1×1×1 convolution layer and a sigmoid function, the attention weight is obtained.

[0029] The source domain features are weighted using attention weights.

[0030] Preferably, the enhanced source domain features are aligned with the target domain features, specifically: using attention consistency loss to constrain the attention map of the source domain image and the attention map of the target domain image, so that the obtained attention map Figure 1 The attention map is generated by the coordinate attention layer during the feature encoder feature extraction process;

[0031] The attention consistency loss is expressed as:

[0032]

[0033] Among them, X and X′ represent the source domain features and target domain features respectively, A l and A l ′ respectively represent the attention maps corresponding to X and X′, represents the spatial deformation from the source domain image to the target domain image; l represents the sequence number of the improved residual block, and L represents the total number of improved residual blocks.

[0034] Preferably, the method of realizing Alzheimer's disease process identification by using the trained Alzheimer's disease process identification model comprises the following steps:

[0035] Preprocessing the brain image data to obtain a preprocessed image as an image to be recognized;

[0036] The image to be identified is input into the trained Alzheimer's disease process identification model, the feature encoder extracts the features of the image to be identified, and the classifier classifies it based on the extracted features to identify the current disease process category.

[0037] The present invention also provides an Alzheimer's disease process identification system based on domain adaptation, comprising:

[0038] The source domain image acquisition module pre-processes the brain image data to obtain the source domain image;

[0039] The target domain image acquisition module performs data enhancement processing on the source domain image to obtain the enhanced source domain image as the target domain image;

[0040] Model training module, which uses target domain images and source domain images to train the Alzheimer's disease process recognition model;

[0041] Model application module, which uses the trained Alzheimer's disease process identification model to realize Alzheimer's disease process identification;

[0042] The Alzheimer's disease process recognition model includes a feature encoder and a classifier. The feature encoder uses a convolutional neural network combined with an attention mechanism to extract features of the image to be recognized, and the classifier classifies and recognizes the process based on the extracted features.

[0043] The present invention has the following beneficial effects:

[0044] (1) The Alzheimer's disease process identification method based on domain adaptation of the present invention can serve as an important indicator to assist in the prediction of Alzheimer's disease, improve doctors' control over the disease and thus formulate diagnostic plans early; at the same time, stable performance can be guaranteed;

[0045] (2) The present invention can effectively standardize brain image data by preprocessing the brain image data; the image data is processed using a data enhancement method, which helps to solve the problem of data scarcity or imbalance and simulates the patient data collected by different hospitals in clinical diagnosis;

[0046] (3) The present invention proposes a domain loss function to constrain the model learning and align the feature expressions of the source domain and target domain features, so that the model learns features with domain-invariant properties, thereby alleviating the impact of differences in data from different sites on the model recognition performance.

[0047] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments, but the present invention is not limited to the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A method step diagram of an embodiment of the present invention;

[0049] Figure 2 A schematic diagram of model training according to an embodiment of the present invention;

[0050] Figure 3 A structural diagram of a coordinate attention module according to an embodiment of the present invention;

[0051] Figure 4 4 is a system structure diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0052] See also Figure 1 FIG. 1 is a diagram showing steps of a method according to an embodiment of the present invention, comprising the following steps:

[0053] S101, preprocessing brain image data to obtain a source domain image;

[0054] S102, performing data enhancement processing on the source domain image to obtain an enhanced source domain image as a target domain image;

[0055] S103, training an Alzheimer's disease progression recognition model using the target domain image and the source domain image;

[0056] S104, using the trained Alzheimer's disease process identification model to implement Alzheimer's disease process identification;

[0057] The Alzheimer's disease process recognition model includes a feature encoder and a classifier. The feature encoder uses a convolutional neural network combined with an attention mechanism to extract features of the image to be recognized, and the classifier classifies and recognizes the process based on the extracted features.

[0058] Specifically, in S101, brain image data is subjected to brain region morphological feature extraction. First, all skulls are stripped from the image, and then the images are registered with each other to segment the main structures of the brain; then, all images are normalized and smoothed to a standard size, and then the volume, cortical thickness, and surface area features of the region of interest are extracted from the Alzheimer's disease image, including the standard deviation of cortical thickness, the average value of cortical thickness, white matter volume, cortical surface area, and cortical volume. Based on the above medical morphological features, this embodiment analyzes the identification ability of each medical morphological feature for the development process of Alzheimer's disease, improves the balance of data distribution, and improves the performance of the algorithm.

[0059] S101 of the embodiment of the present invention specifically includes: performing head motion correction, non-uniform intensity correction, Talairach space transformation, standardization and skull stripping on the brain image data; and realigning all data to the MNI space using SPM12 software.

[0060] Specifically, in S102, a spatial deformation data enhancement method is used, which specifically includes flipping of the axial plane and random changes in scale.

[0061] Specifically, in S103, the training process of the Alzheimer's disease process recognition model is shown in Figure 2 As shown, where f s and f t They represent the source domain features and target domain features of the source domain image and the target domain image after the feature encoder and global average pooling operation respectively; w c is the weight value corresponding to the category c of the source domain image, and the weight value w c and source domain features f s Multiply to get f s p

[0062] The training process consists of the following steps:

[0063] S1031, inputting the source domain image into a feature encoder to obtain source domain image features;

[0064] S1032, inputting the target domain image into a feature encoder to obtain target domain image features;

[0065] S1033, using source domain image features and target domain image features for adversarial learning, so that the feature encoder learns features with domain-invariant properties;

[0066] S1034, using the weight information from the classifier to weight the source domain image features to enhance the expression of the source domain features, aligning the enhanced source domain features with the target domain features, and improving the classifier's perception of pathological details.

[0067] Specifically, the feature encoder includes three stacked improved residual blocks, each residual block includes a first 3×3×3 convolution layer, a second 3×3×3 convolution layer, a first coordinate attention layer, a third 3×3×3 convolution layer, a fourth 3×3×3 convolution layer and a second coordinate attention layer connected in sequence; the outputs of the three improved residual blocks pass through a global average pooling layer together, and the output of the global average pooling layer is used as the extracted feature.

[0068] Specifically, the feature encoder extracts features of the image to be identified, and the image to be identified first passes through a 7×7×7 convolution layer and a 3×3×3 maximum pooling layer in sequence before entering the feature encoder.

[0069] Specifically, S1033 proposes a domain loss function, uses a classifier to build a domain discriminator, inputs the source domain image features and the target domain image features as two categories into the domain discriminator, and maximizes the domain loss function so that the discriminator cannot determine whether the data is from the source domain or the target domain, thereby narrowing the gap between the source domain and target domain features, and allowing the feature encoder to learn features with domain-invariant properties.

[0070] Specifically, the domain loss function is expressed as:

[0071]

[0072] Among them, i represents the data sequence number, N represents the number of samples; y i D Represents data x i The field label, y i =1 indicates data x i From the source domain, y i =0 means data x i from the target domain; F represents the feature encoder, C D Represents the domain discriminator.

[0073] Specifically, in S1034, the step of weighting the source domain image features by using the weight information from the classifier to strengthen the expression of the source domain features includes the following steps:

[0074] Three pooling layers with filter sizes of (H, W, 1), (1, W, D) and (H, 1, D) are used to pool each channel of the input to obtain a feature map that can perceive direction, where H, W and D are the height, width and depth of the source domain features respectively;

[0075] The obtained feature maps are concatenated in the same spatial dimension, input into a 1×1×1 convolution layer, BN normalization and ReLU activation function in sequence, and then split into three vectors. After passing through a 1×1×1 convolution layer and a sigmoid function, the attention weight is obtained.

[0076] The source domain features are weighted using attention weights.

[0077] Specifically, in S1034, the enhanced source domain features are aligned with the target domain features, specifically: using the attention consistency loss to constrain the attention map of the source domain image and the attention map of the target domain image, so that the obtained attention map Figure 1The attention map is generated by the coordinate attention layer in the process of feature encoder extracting features (the feature encoder generates the attention map corresponding to the source domain features in the process of feature encoder extracting features from the source domain image; the feature encoder generates the attention map corresponding to the target domain features in the process of feature encoder extracting features from the target domain image); the attention consistency loss is expressed as:

[0078]

[0079] Among them, X and X′ represent the source domain features and target domain features respectively, A l and A l ′ respectively represent the attention maps corresponding to X and X′, represents the spatial deformation from the source domain image to the target domain image; l represents the sequence number of the improved residual block, and L represents the total number of improved residual blocks.

[0080] Specifically, the S104 includes the following steps:

[0081] S1041, preprocessing the brain image data to obtain a preprocessed image as an image to be recognized;

[0082] S1042, input the image to be identified into the trained Alzheimer's disease progression identification model, the feature encoder extracts the features of the image to be identified, and the classifier classifies based on the extracted features to identify the current disease progression category.

[0083] See also Figure 4 As shown, the Alzheimer's disease process identification system based on domain adaptation according to an embodiment of the present invention includes:

[0084] The source domain image acquisition module 401 pre-processes the brain image data to obtain a source domain image;

[0085] The target domain image acquisition module 402 performs data enhancement processing on the source domain image to obtain an enhanced source domain image as the target domain image;

[0086] A model training module 403, using the target domain image and the source domain image to train an Alzheimer's disease progression recognition model;

[0087] A model application module 404 implements Alzheimer's disease process identification using the trained Alzheimer's disease process identification model;

[0088] The Alzheimer's disease process recognition model includes a feature encoder and a classifier. The feature encoder uses a convolutional neural network combined with an attention mechanism to extract features of the image to be recognized, and the classifier classifies and recognizes the process based on the extracted features.

[0089] The embodiment of the present invention can effectively standardize brain image data by preprocessing the brain image data; use data enhancement methods to process image data, which helps solve the problem of data scarcity or imbalance, and simulates patient data collected by different hospitals in clinical diagnosis; by proposing a domain loss function to constrain the model's learning and aligning the feature expression of source domain and target domain features, the model learns features with domain-invariant properties, alleviating the impact of data differences at different sites on the model's diagnostic performance; it can serve as an important indicator to assist in the prediction of Alzheimer's disease, improve doctors' control over the disease, and thus formulate diagnostic plans early. At the same time, the present invention can ensure stable performance.

[0090] Although the specific implementation modes of the present invention are described above, those skilled in the art should understand that the specific implementation modes described are only illustrative and are not intended to limit the scope of the present invention. Equivalent modifications and changes made by those skilled in the art in accordance with the spirit of the present invention should be included in the scope of protection of the claims of the present invention.

Claims

1. A method for identifying Alzheimer's disease process based on domain adaptation, characterized in that: The steps include: Preprocess the brain image data to obtain source domain images; Perform data enhancement processing on the source domain image to obtain the enhanced source domain image as the target domain image; Using target domain images and source domain images to train an Alzheimer's disease progression recognition model; Using the trained Alzheimer's disease process identification model to achieve Alzheimer's disease process identification; The Alzheimer's disease process recognition model includes a feature encoder and a classifier. The feature encoder uses a convolutional neural network combined with an attention mechanism to extract features of the image to be recognized, and the classifier classifies and recognizes the process based on the extracted features.

2. The method for identifying Alzheimer's disease process based on domain adaptation according to claim 1, characterized in that: The preprocessing of brain image data includes: The brain imaging data were corrected for head motion, non-uniform intensity, Talairach space transformation, normalization, and skull stripping; All data were realigned to MNI space using SPM12 software.

3. The method for identifying Alzheimer's disease process based on domain adaptation according to claim 1, characterized in that: The method of training the Alzheimer's disease progression recognition model using the target domain image and the source domain image includes: Input the source domain image into the feature encoder to obtain the source domain image features; Input the target domain image into the feature encoder to obtain the target domain image features; Using source domain image features and target domain image features for adversarial learning, the feature encoder learns features with domain-invariant properties. The weight information from the classifier is used to weight the source domain image features to enhance the expression of the source domain features, and the enhanced source domain features are aligned with the target domain features to improve the classifier's perception of pathological details.

4. The method for identifying Alzheimer's disease process based on domain adaptation according to claim 3, characterized in that: The feature encoder includes three improved residual blocks and a global average pooling layer, each improved residual block performs feature encoding on the input image to be recognized, the outputs of the three improved residual blocks are input into the global average pooling layer together, and the output of the global average pooling layer is used as the extracted feature; each improved residual block includes a first 3×3×3 convolutional layer, a second 3×3×3 convolutional layer, a first coordinate attention layer, a third 3×3×3 convolutional layer, a fourth 3×3×3 convolutional layer and a second coordinate attention layer connected in sequence.

5. The method for identifying Alzheimer's disease process based on domain adaptation according to claim 4, characterized in that: The feature encoder extracts features of the image to be identified. Before entering the feature encoder, the image to be identified first passes through a 7×7×7 convolution layer and a 3×3×3 maximum pooling layer in sequence.

6. The method for identifying Alzheimer's disease process based on domain adaptation according to claim 1, characterized in that: The method of using source domain image features and target domain image features for adversarial learning includes the following steps: A classifier is used to build a domain discriminator, and the source domain image features and the target domain image features are input into the domain discriminator as two categories; A domain loss function is proposed, expressed as: Among them, i represents the data sequence number, and N represents the number of data; Represents data x i The field label, y i =1 indicates data x i From the source domain, y i =0 means data x i from the target domain; F represents the feature encoder, C D represents the domain discriminator; By maximizing the domain loss function, the domain discriminator cannot determine whether the data comes from the source domain or the target domain, thereby narrowing the gap between the source domain features and the target domain features, and allowing the feature encoder to learn features with domain-invariant properties.

7. The method for identifying Alzheimer's disease process based on domain adaptation according to claim 3, characterized in that: The step of weighting the source domain image features by using the weight information from the classifier to strengthen the expression of the source domain features comprises the following steps: Three pooling layers with filter sizes of (H, W, 1), (1, W, D) and (H, 1, D) are used to pool each channel of the source domain features to obtain feature maps that can perceive direction, where H, W and D are the height, width and depth of the source domain features respectively; The obtained feature maps are concatenated in the same spatial dimension, input into a 1×1×1 convolution layer, BN normalization and ReLU activation function in sequence, and then split into three vectors. After passing through a 1×1×1 convolution layer and a sigmoid function, the attention weight is obtained. The source domain features are weighted using attention weights.

8. The method for identifying Alzheimer's disease process based on domain adaptation according to claim 4, characterized in that: The enhanced source domain features are aligned with the target domain features, specifically: using attention consistency loss to constrain the attention map of the source domain image and the attention map of the target domain image, so that the obtained attention maps are consistent; the attention map is generated by the coordinate attention layer in the feature encoder feature extraction process; The attention consistency loss is expressed as: Among them, X and X′ represent the source domain features and target domain features respectively, A l and A l ′ respectively represent the attention maps corresponding to X and X′, represents the spatial deformation from the source domain image to the target domain image; l represents the sequence number of the improved residual block, and L represents the total number of improved residual blocks.

9. The method for identifying Alzheimer's disease process based on domain adaptation according to claim 1, characterized in that: The method of using the trained Alzheimer's disease process identification model to implement Alzheimer's disease process identification includes the following steps: Preprocessing the brain image data to obtain a preprocessed image as an image to be recognized; The image to be identified is input into the trained Alzheimer's disease process identification model, the feature encoder extracts the features of the image to be identified, and the classifier classifies it based on the extracted features to identify the current disease process category.

10. An Alzheimer's disease process identification system based on domain adaptation, characterized in that: include: The source domain image acquisition module pre-processes the brain image data to obtain the source domain image; The target domain image acquisition module performs data enhancement processing on the source domain image to obtain the enhanced source domain image as the target domain image; Model training module, which uses target domain images and source domain images to train the Alzheimer's disease process recognition model; Model application module, which uses the trained Alzheimer's disease process identification model to realize Alzheimer's disease process identification; The Alzheimer's disease process recognition model includes a feature encoder and a classifier. The feature encoder uses a convolutional neural network combined with an attention mechanism to extract features of the image to be recognized, and the classifier classifies and recognizes the process based on the extracted features.