An alzheimer's disease classification method based on specific brain region multi-relation reasoning

Abstract

For the auxiliary diagnosis task of Alzheimer's disease, an Alzheimer's disease classification method based on specific brain region multi-relation reasoning is disclosed. Fully learning the multi-relation perception representation of the disease-related region in sMRI image, including spatial relationship and topological information, is the key to improve the accuracy of auxiliary diagnosis. The invention regards the distinction of disease state as a graph classification problem, and constructs a spatial graph and a semantic graph. A hollow convolution module is designed to learn specific region representation, and a multi-relation graph convolution module is used to capture various types of brain region relationship. Global reasoning is performed on the learned graph structure to select discriminative information and generate global representation for diagnosis. Based on the advantages of deep learning and the characteristics of sMRI image, the Alzheimer's disease classification network based on specific brain region multi-relation reasoning is designed, which has broad application prospects in medical image analysis, Alzheimer's disease auxiliary diagnosis and other aspects.

Description

Technical Field

[0001] This invention designs an Alzheimer's disease classification method based on multi-relational reasoning of specific brain regions. It realizes multi-relational perceptual representation of disease-related regions in sMRI images and performs global reasoning on the multi-relational region representation to improve the diagnostic accuracy of Alzheimer's disease. It involves deep learning technology, medical image processing technology and other fields. Background Technology

[0002] Alzheimer's disease (AD) is a neurodegenerative and irreversible brain disease, and the most common cause of dementia. Currently in China, approximately 15 million people over the age of 60 suffer from Alzheimer's disease, while about 9.83 million suffer from Alzheimer's disease. Alzheimer's disease begins with memory impairment, and as the disease progresses, problems arise in communication and body control, making it one of the leading causes of death in the elderly. To date, there is no cure for Alzheimer's disease, but symptoms can be alleviated or its progression slowed through medication, exercise, and memory training. Therefore, accurate diagnosis of Alzheimer's disease is crucial, helping to slow disease progression and improve the patient's overall health. With the development of medical imaging technology, brain imaging has become an important basis for the diagnosis and assessment of Alzheimer's disease progression. Structural magnetic resonance imaging (sMRI), as a non-invasive imaging technique, can generate detailed three-dimensional anatomical images of the brain, simulating the anatomical changes in the brain affected by the disease. Currently, various deep learning methods based on sMRI images have been used to identify abnormal structural changes in the brain and explore disease-related imaging biomarkers, proving their effectiveness. However, most of these methods extract features from whole-brain images or simply combine features extracted from local image patches, neglecting the topological relationships between brain regions. The spatial relationships and structural information between brain regions in sMRI images of different disease states will inevitably differ. Furthermore, Alzheimer's disease affects different brain regions differently, and using whole-brain sMRI data for diagnosis may contain irrelevant redundant information, leading to decreased diagnostic performance. To address this, computer-aided diagnostic methods based on sMRI images can effectively improve AD diagnosis performance by learning local brain features and relationships between brain regions. How to utilize computer vision technology to select disease-related brain regions and extract discriminative features from these regions to improve the accuracy of assisted diagnosis has been a research hotspot in recent years. Summary of the Invention

[0003] This invention, leveraging the characteristics and advantages of sMRI imaging and deep learning, provides a method for Alzheimer's disease classification based on multi-relational reasoning of specific brain regions. A key region extraction module is proposed, using Kendall correlation coefficients for inter-group difference analysis to identify specific brain regions strongly correlated with the disease. A region feature extraction module is constructed, utilizing dilated convolution to learn more information representations of the discriminant regions, including near and far-range information. A multi-relational reasoning network for specific brain regions is designed, employing graph convolution to encode spatial and semantic relationships in the constructed spatial and semantic graphs, enriching the regional-level representation for Alzheimer's disease diagnosis. Based on the multi-relational perceived regional features, global reasoning is performed to obtain the final feature representation for classification, which is then fed into two fully connected layers and a softmax activation function to obtain the final diagnostic result. Utilizing artificial intelligence technology to diagnose Alzheimer's disease from sMRI images can assist doctors in making more timely and accurate judgments about patients and potential patients, improving the stability and reliability of screening and diagnosis. Therefore, research on Alzheimer's disease auxiliary diagnosis based on sMRI images is particularly important.

[0004] The data used in this invention comes from the authoritative Alzheimer's disease public dataset ADNI. We selected 3492 high-quality sMRI images from 1650 subjects in the ADNI database as experimental data and formulated a data leakage-free partitioning rule. Subjects were divided into three categories: cognitively unified (CN), mild cognitive impairment (MCI), and Alzheimer's disease (AD). Our method improved diagnostic performance on the ADNI dataset, and the rationality of the model was verified through ablation experiments.

[0005] The present invention achieves the above objectives through the following technical solutions:

[0006] 1. In step one, a dataset is constructed. For Alzheimer's disease research, the largest publicly available dataset is ADNI. The dataset in this invention uses 3492 T1-weighted sMRI images from 1650 subjects in the ADNI database. To avoid data leakage, the training and test sets are divided by person ID; simultaneously, to achieve a balanced sample distribution, the training and test sets have minimal differences in clinical scale scores and demographic information. The sMRI images are preprocessed using the Clinica software platform. First, tissue segmentation, registration, and spatial standardization are performed; then, intensity normalization is performed using max-min normalization; furthermore, the internationally recognized standard brain template AAL3 is selected for brain region segmentation; and the average gray matter density of 166 anatomical regions obtained from brain atlases is calculated.

[0007] 2. In step two, this invention proposes a key region extraction method based on inter-group difference analysis. Structural changes in certain brain regions are strongly correlated with diseases. Therefore, based on local morphological characteristics, we rank the importance of information regions through group comparison to identify brain regions with significant inter-group differences. Kendall correlation analysis was performed on the gray matter density of each brain region and disease category in the training set. Statistical analysis was performed using SPSS 26, and the absolute value of the Kendall correlation coefficient represents the difference between each brain region in the three disease states; the larger the absolute value, the stronger the specificity of the region. The top six regions were selected for experimental analysis, including the bilateral hippocampus, bilateral amygdala, and bilateral parahippocampal gyrus. These regions are consistent with medical research, with clinical manifestations mainly characterized by histopathological changes.

[0008] 3. In step three, this invention proposes a feature extraction network based on dilated convolution. Dilated convolutions with dilation coefficients of 1, 2, and 5, and kernels of 3×3×3, are used to learn more information representations of the discriminative regions, including near-range and far-range information. Finally, global average pooling is used to generate region features. The use of dilated convolutions allows the model to extract features from different and relatively low-resolution brain regions without additional computational overhead.

[0009] 4. In step four, this invention proposes a specific brain region multi-relation reasoning network. First, a spatial graph and a semantic graph are constructed, where nodes represent the features of specific regions, and edges represent the semantic or spatial relationships between them. Semantic relationship encoding uses dynamic graph convolution to adaptively obtain the semantic relationship perception between brain regions, enriching the regional-level representation of Alzheimer's disease diagnosis, as shown in formula (1). Since the spatial graph is directional and contains label information, labels for different directions and edges need to be transformed using separate transformation matrices and biases. Therefore, spatial relationship encoding uses graph convolution with attention mechanism to reason about the spatial graph, which is sensitive to the directional aggregation information from neighboring nodes and automatically pays attention to potentially important edges, as shown in formula (2).

[0010]

[0011] Where δ is the sigmoid activation function and f is the LeakyReLU activation function. For node features, W1 and W r This represents the weight matrix.

[0012]

[0013] Where W is the weight matrix, σ is a nonlinear function, and N(v) i ) represents the set of adjacent nodes, α ij This represents the attention coefficient.

[0014] 5. In step five, this invention proposes a classification process based on global reasoning. First, the region features based on semantic perception and spatial perception are added together to obtain the final region-level representation. Then, this representation is sequentially fed into a gate-controlled recurrent unit for global reasoning, selecting discriminative information and filtering out unimportant information to generate the final representation for classification. Finally, two fully connected layers and a softmax activation function are used to predict the probability score of the input sMRI image belonging to a specific category. Cross-entropy is used as the loss function to evaluate the difference between the predicted label and the true label.

[0015] The main content of this invention is to propose a classification research method for Alzheimer's disease based on multi-relational reasoning of specific brain regions. The designed Alzheimer's disease classification network can focus on the spatial relationships and topological information of specific regions in sMRI images, improving the accuracy of diagnosis and having significant implications for medical image analysis and auxiliary diagnosis of Alzheimer's disease. Attached Figure Description

[0016] Figure 1 This is a flowchart of the preprocessing process for sMRI images.

[0017] Figure 2 This is a structural diagram of the proposed Alzheimer's disease classification algorithm based on multi-relational reasoning of specific brain regions. Detailed Implementation

[0018] The present invention will be further described below with reference to the accompanying drawings:

[0019] Figure 1 The flowchart shows the preprocessing process for sMRI images. First, the sMRI images are converted to BIDS format, and tissue segmentation, registration, and spatial normalization are performed. Then, the pixel values ​​are normalized to [0, 1] using max-min normalization. Finally, the sMRI scans are segmented into brain regions using the standard brain template AAL3, and the average gray matter density of each anatomical region is calculated.

[0020] Figure 2This invention presents an Alzheimer's disease classification method based on multi-relational reasoning of specific brain regions. Specific regions in sMRI images are extracted through inter-group difference analysis, and rich regional feature representations are further extracted using a feature extraction network based on dilated convolution. To learn the spatial location relationships and topological information of brain regions, a multi-relational reasoning network based on graph convolution is used to reason about semantic and spatial relationships, enriching the regional-level representation for Alzheimer's disease diagnosis. To achieve effective aggregation of multi-relational region features, global reasoning is used to filter out unimportant information, generating the final feature representation for classification. Ablation experiments were conducted to verify the effectiveness of the proposed method. For the AD and CN binary classification tasks, multiple methods using whole-brain images and different numbers of paired brain regions were trained under the same experimental environment, and the test results are shown in Table 1. Simultaneously, to verify the rationality of the specific brain region multi-relational reasoning network module proposed in this invention, ablation experiments were conducted, experimentally verifying methods using only spatial reasoning or semantic reasoning and methods considering both reasoning simultaneously, and the results are shown in Table 2.

[0021] Table 1

[0022]

[0023] Table 2

[0024]

[0025] To verify the effectiveness of the method proposed in this invention, several Alzheimer's disease diagnostic models based on sMRI images and classic image classification models ResNet18, ResNet50, and ResNet101 were selected for comparison with the method proposed in this invention. The symbol * indicates that the experimental data and data preprocessing procedures used in this method differ from those in this invention. The test results are shown in Table 3.

[0026] Table 3

[0027]

[0028] As shown in Table 3, the method proposed in this invention achieves further performance improvements over other models in Alzheimer's disease classification based on sMRI images. Therefore, the effectiveness of the method proposed in this invention is verified.

Claims

1. A method for classifying Alzheimer's disease based on multi-relational reasoning of specific brain regions, characterized in that... Includes the following steps: Step 1: Based on the publicly available ADNI dataset, select high-quality sMRI images to form the experimental data, and formulate non-leaking data partitioning criteria to divide the data into training and test sets with balanced sample distribution; perform tissue segmentation, registration and spatial normalization preprocessing on the sMRI images in the dataset, and then perform brain region segmentation based on the standard brain template AAL3 and calculate the average gray matter density of each brain region. Step 2: Extraction of key brain regions. Calculate the Kendall correlation coefficient between the brain regions and disease categories obtained in Step 1. After intergroup difference analysis, identify brain regions with significant intergroup differences and select the top 6 regions for experimental analysis. Step 3: Based on the specific regions obtained in Step 2, a rich region feature representation, including near-range information and far-range information, is extracted using a feature extraction network based on dilated convolution. Step 4: Propose a brain region-specific multi-relation reasoning network, and encode the constructed spatial graph and semantic graph using graph convolution to enrich the regional representation for Alzheimer's disease diagnosis; Step 5: Introduce global inference for aggregating features of multi-relational perceptual regions to obtain the final feature representation for classification, and optimize the prediction results using cross-entropy loss.

2. The Alzheimer's disease classification method based on multi-relational reasoning of specific brain regions according to claim 1, characterized in that... The dataset creation and preprocessing process in Step 1 is as follows: Based on the publicly available ADNI dataset, high-quality sMRI images constitute the experimental data used, and non-disclosure data partitioning criteria are established to divide the training and test sets, while considering the statistical parameters of the samples, including the balanced distribution of clinical scale scores and demographic information; the sMRI images are preprocessed, including tissue segmentation, registration, spatial standardization, and intensity normalization; and the internationally recognized standard brain template AAL3 is selected for brain region segmentation, and the average gray matter density of each anatomical region is calculated.

3. The Alzheimer's disease classification method based on multi-relational reasoning of specific brain regions according to claim 1, characterized in that... In step two, the importance of anatomical regions was ranked based on local morphological features through intergroup difference analysis to identify brain regions with significant intergroup differences. Kendall correlation coefficient was used to represent the differences of each brain region in the three disease states. The larger the absolute value, the stronger the specificity of the region. The top 6 brain regions were selected for experimental analysis.

4. The Alzheimer's disease classification method based on multi-relational reasoning of specific brain regions according to claim 1, characterized in that... In step three, features are extracted from the anatomical regions obtained in step two. Dilated convolutions with dilation coefficients of 1, 2, and 5 are used to learn more information representations of the discriminative regions, including near-field and far-field information. Finally, regional features are generated through global average pooling. Dilated convolutions allow the model to extract features from different and relatively low-resolution brain regions without additional computational overhead.

5. The Alzheimer's disease classification method based on multi-relational reasoning of specific brain regions according to claim 1, characterized in that... In step four, multi-relation reasoning is performed on specific brain regions. First, spatial graphs and semantic graphs are constructed, where nodes represent the features of specific regions and edges represent the semantic or spatial relationships between them. Then, a graph convolutional network is used to learn the multi-relationships of specific regions, including spatial relationships and topological information. Specifically, dynamic graph convolution is used to adaptively obtain the semantic relationships between brain regions, and graph convolution with attention mechanism is used to reason about the spatial graph to obtain spatial relationships, thereby aggregating the information of adjacent nodes and automatically paying attention to potentially important edges.

6. The Alzheimer's disease classification method based on multi-relational reasoning of specific brain regions according to claim 1, characterized in that... In step five, global reasoning is performed on the multi-relation region representations to select discriminative information, filter out unimportant information, and generate the final feature representation for classification, thereby further improving the accuracy of Alzheimer's disease auxiliary diagnosis based on sMRI images.

Images