Alzheimer's disease diagnosis system based on graph attention and hierarchical encoding

Abstract

The application discloses an Alzheimer's disease diagnosis system based on graph attention and hierarchical coding, relates to the field of neuroimaging data analysis, and comprises a brain region feature sequence construction module, a brain region feature sequence coding module, a brain region association modeling module, a global encoder and a classification prediction module.

Description

Technical Field

[0001] This invention relates to the field of neuroimaging data analysis, and more particularly to an Alzheimer's disease diagnostic system based on graph attention and hierarchical coding. Background Technology

[0002] The statements in this section are merely background information relating to this disclosure and do not necessarily constitute prior art.

[0003] Alzheimer's disease (AD) is a progressive neurodegenerative disease, and early diagnosis is crucial for slowing disease progression and improving patients' quality of life.

[0004] Brain structure analysis based on magnetic resonance imaging (MRI) has become an important non-invasive method for assisting in the diagnosis of Alzheimer's disease (AD). In existing technologies, extracting quantitative features from structural MRI images, such as brain region volume and cortical thickness, and combining them with machine learning or deep learning models for classification is a common approach to disease diagnosis.

[0005] In existing technologies, voxel-based morphometry (VBM) methods distinguish AD patients from normal controls by analyzing changes in gray matter density or volume at the whole-brain voxel level. However, VBM methods suffer from excessively high feature dimensionality, leading to the "curse of dimensionality" and overfitting. Furthermore, they analyze the features of each voxel in isolation, ignoring the inherent spatial correlations between voxels. Surface-based morphometry (SBM) methods can extract more neurobiologically significant features such as cortical thickness and surface area, and often combine atlases to divide the brain into several regions of interest (ROIs) to reduce feature dimensionality. However, traditional machine learning models (such as support vector machines) treat these brain region features as independent input variables, failing to fully utilize the inherent anatomical adjacency relationships and continuous patterns of structural changes between brain regions.

[0006] However, the existing solutions still have the following unresolved shortcomings: First, they treat the morphological features of each brain region as independent input variables, completely severing the inherent anatomical adjacency relationships and structural covariation patterns between brain regions. In reality, the pathological changes in Alzheimer's disease (AD) do not occur randomly and independently, but rather exhibit continuous and progressive spatial propagation along specific anatomical connectivity pathways. Existing solutions fail to fully utilize the inherent anatomical adjacency relationships and continuous patterns of structural changes between brain regions. Second, due to the aforementioned independence assumption, existing models lack information complementarity and smoothing constraints between features of adjacent brain regions, which can easily lead to overfitting and decreased generalization ability in classification decisions, and cannot effectively screen out key brain region clusters with disease discriminative power to assist doctors in clinical diagnosis. Summary of the Invention

[0007] To overcome the shortcomings of the prior art, this invention provides an Alzheimer's disease diagnostic system based on graph attention and hierarchical coding. It aims to diagnose Alzheimer's disease by utilizing cortical thickness extracted from magnetic resonance images and to screen out the brain regions where lesions occur, thereby assisting doctors in diagnosis and reducing their diagnostic burden.

[0008] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: In a first aspect, the present invention provides an Alzheimer's disease diagnostic system based on graph attention and hierarchical coding, comprising: The image preprocessing module is configured to: acquire magnetic resonance imaging data of the object to be analyzed, and extract the cortical thickness features corresponding to each brain region after brain region division; The brain region feature sequence construction module is configured to: construct brain region feature sequences and adjacency matrices based on the cortical thickness features corresponding to each brain region and the anatomical adjacency relationships between brain regions; encode the brain region feature sequences based on a graph attention network, and introduce a graph attention mechanism in conjunction with the adjacency matrix to obtain updated brain region feature sequences. The brain region association modeling module is configured to: divide all brain regions into multiple brain region groups, obtain corresponding intra-group feature representations, input the intra-group feature representations of each brain region group into the corresponding local encoder for modeling, and obtain group-level representations; input the group-level representations into the global encoder for cross-group modeling, and obtain global brain structure representations. The classification prediction module is configured to input the global brain structure representation into the classification layer for prediction and obtain disease prediction results.

[0009] A further technical solution involves constructing a degree matrix based on the adjacency matrix, and then normalizing the adjacency matrix based on the degree matrix to obtain a normalized adjacency matrix.

[0010] Further technical solutions yield updated brain region feature sequences, specifically including: The brain region feature sequences are mapped to a high-dimensional space to obtain the initial hidden representation of the brain region; By combining the adjacency matrix, the attention weights of each brain region and its neighboring brain regions are calculated based on the initial hidden representation of the brain region. After normalization, the normalized attention weights are obtained. Based on the normalized attention weights, information aggregation under topological constraints is performed on the brain region feature sequences to obtain the updated brain region feature sequences.

[0011] A further technical solution is to divide all brain regions into multiple brain region groups based on their anatomical affiliation.

[0012] A further technical solution yields group-level representations, specifically including: inputting the feature representations within each brain region group into the corresponding local encoder for modeling to obtain the local feature representations of all brain regions within each brain region group; and pooling and aggregating the local feature representations of each brain region group to obtain group-level representations.

[0013] A further technical solution yields a global brain structure representation, specifically including: combining all group-level representations in a preset order and inputting them into a global encoder for cross-group modeling.

[0014] A further technical solution involves using a joint loss function to train the model end-to-end, as follows:

[0015] in, Denotes the joint loss function. , Represents the balance coefficient. Represents cross-entropy loss, This represents the graph smoothing loss. This represents the group-level consistency constraint loss.

[0016] A further technical solution is that the group-level consistency constraint loss is expressed as:

[0017] in, This represents the group-level consistency constraint loss. Indicates the first Group-level representation of brain regions This represents the average characteristic of all group-level representations.

[0018] In a second aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the contents of the Alzheimer's disease diagnostic system based on graph attention and hierarchical coding as described in the first aspect.

[0019] Thirdly, the present invention provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the contents of the Alzheimer's disease diagnostic system based on graph attention and hierarchical coding as described in the first aspect.

[0020] The above one or more technical solutions have the following beneficial effects: This invention utilizes graph attention networks combined with anatomical adjacency relationships between brain regions to adaptively aggregate information on cortical thickness features of each brain region under topological constraints. This effectively captures the spatial propagation patterns of Alzheimer's disease pathological changes between adjacent brain regions. Simultaneously, by dividing all brain regions into multiple brain region groups and sequentially employing local encoders and global encoders for fine-grained intra-group modeling and cross-group modeling, multi-level feature fusion from fine-grained brain regions to coarse-grained functional groups is achieved. This results in a more discriminative and robust global brain structure representation, ultimately significantly improving the accuracy of disease prediction.

[0021] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0022] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0023] Figure 1 This is a flowchart of the Alzheimer's disease diagnostic system based on graph attention and hierarchical coding according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the diagnostic model according to an embodiment of the present invention; Figure 3 This is the confusion matrix in the AD, MCI, and NC classification tasks of this invention. Figure 4 This is the confusion matrix in the MCI and NC classification tasks of this invention embodiment; Figure 5 This is the confusion matrix in the AD and NC classification tasks of this invention embodiment; Figure 6 This is the confusion matrix in the AD and MCI classification tasks of this invention embodiment; Figure 7 This is a ranking of the contribution levels of brain regions in the AD, MCI, and NC classifications of this invention. Figure 8 This is a ranking of the contribution levels of brain regions in the MCI and NC classifications of this invention. Figure 9 This is a ranking of the contribution levels of brain regions in the AD and NC classifications of this invention. Figure 10 This is a ranking of the contribution levels of brain regions in the AD and MCI classifications of this invention. Detailed Implementation

[0024] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0025] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0026] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0027] Example 1 The biological characteristics of Alzheimer's disease mainly include: -Amyloid protein ( Senile plaques, neuronal cell loss, and neurofibrillary tangles accumulate in the cerebral cortex and hippocampus of Alzheimer's patients. Neurofibrillary tangles are composed of highly phosphorylated Tau protein, so the Tau protein content in Alzheimer's patients is generally higher than in normal individuals. Highly phosphorylated Tau protein damages the microtubule structure of neurons, leading to functional impairment and ultimately causing Alzheimer's disease.

[0028] In neuroimaging, clinical diagnosis of Alzheimer's disease (AD) primarily utilizes magnetic resonance imaging (MRI) and computed tomography (CT). MRI is the preferred examination due to its clearer imaging of brain tissue, and commonly used sequences include T1-weighted imaging, T2-weighted imaging, fluid-attenuated inversion recovery imaging (horizontal + hippocampal coronal views), and diffusion-weighted imaging. Doctors observe changes in the patient's brain on MRI images, such as the degree of hippocampal atrophy and ventricular enlargement, using these potential changes to make a diagnosis. However, this diagnostic method is empirical, and the results are dependent on factors such as hospital equipment and the doctor's experience, and cannot guarantee an accurate diagnosis. For patients who cannot undergo cranial MRI due to reasons such as metallic foreign bodies or claustrophobia, multi-slice CT scans are recommended as an alternative. Currently, resting-state and task-based functional MRI are only used for clinical research on the brain network mechanisms of AD and cannot yet be truly used for the clinical diagnosis of AD.

[0029] Molecular imaging methods: Positron emission tomography (PET) plays a crucial role in the diagnosis of Alzheimer's disease (AD) throughout its entire course, from the preclinical stage to the dementia stage. Depending on the labeled substances, AD-related PET examinations mainly include... -Amyloid -protein, PET-PET, Tau-PET, and fluorodeoxyglucose (FDG)-PET are all options. However, PET scans are expensive and have drawbacks. Only a few hospitals have the capability to perform PET and Tau-PET scans, thus limiting the clinical application of this examination.

[0030] Humoral markers for Alzheimer's disease (AD) come from three sources: cerebrospinal fluid, blood, and urine. These mainly include those mentioned in the pathogenesis section above. Related biomarkers, Tau protein-related biomarkers.

[0031] In the field of computer science, machine learning and deep learning methods excel at capturing subtle changes. Therefore, research on computer-aided machine learning to predict the progression of Alzheimer's disease has received considerable attention.

[0032] Research on neuroimaging data has attracted widespread attention. Well-known machine learning methods, such as Linear Discriminant Analysis (LDA), Decision Tree (DT), and Support Vector Machines (SVM), have been adopted and hold promise for early diagnosis and prediction of Alzheimer's disease (AD) progression. However, appropriate preprocessing steps must be applied before using such methods. Furthermore, these methods require feature extraction, feature selection, dimensionality reduction, and feature-based classification for classification and prediction.

[0033] In current research, preprocessing for magnetic resonance imaging (MRI) is largely based on voxel-based morphometry (VBM). In voxel-level methods, tissue features (e.g., gray matter density) extracted from structural magnetic resonance imaging (sMRI) scans constitute high-dimensional voxel structural features for AD diagnosis. For example, existing techniques combine the voxel sizes of both gray matter (GM) and white matter (WM) in the brain and then use PCA for feature extraction to diagnose Alzheimer's disease.

[0034] However, the number of training images used for AD classification is too small compared to the dimension of the features, often leading to the curse of dimensionality (a phenomenon where computational complexity increases exponentially with increasing dimension). Voxel-level methods aim to distinguish disease-related microstructures in MRI images of patients and normal controls, using voxel and tissue (e.g., gray and white matter) density as features for classification algorithms. However, analyzing features only on isolated voxels leads to ignorance of the high correlations between voxels. Another limitation of VBM is the overfitting problem, as the feature representations extracted using VBM consistently show high discrepancies compared to the small number of subjects used for model training.

[0035] Surface-based morphology (SBM) is another method for extracting features from sMRI. While voxel-based morphology (VBM) quantitatively measures the density and volume of gray and white matter through inter-voxel comparisons, SBM can measure cortical thickness, surface area, concavity, curvature, and other factors beyond volume—its most significant advantage over VBM. The cerebral cortex, the highest-level neural center in the brain, consists of gray matter arranged in a specific order on the brain's surface. Therefore, features such as curvature, surface area, and cortical thickness can more accurately describe the arrangement of various nerve cells and fibers. For example, existing techniques extract volumetric features from 93 regions of interest (ROIs) automatically labeled by an atlas warping algorithm and use linear support vector machines (SVMs) for AD classification. Existing techniques also segment hippocampal features from sMRI scans for AD diagnosis and MCI conversion prediction.

[0036] The application of deep learning in Alzheimer's disease has evolved from initial Convolutional Neural Networks (CNNs) to subsequent introductions of attention mechanisms or Transformer structures. For example, existing technologies have proposed a dual-attention multi-instance deep learning network, including a spatial attention module and multi-instance attention, for the early diagnosis of Alzheimer's disease and its prodromal mild cognitive impairment. Another existing technology proposes a hybrid network with multi-scale attention convolution and an age Transformer for AD diagnosis. This network first extracts image features using multi-scale attention convolution, then embeds age information into these features, and finally inputs this information into the age Transformer network to capture the dependencies between subjects of different ages.

[0037] The technical problem addressed by this invention is how to model the anatomical adjacency relationships and continuous structural changes of brain regions without significantly increasing feature dimensions, and to achieve interpretable screening of lesion regions. To address the above technical problem, this invention proposes an Alzheimer's disease diagnostic system based on graph attention and hierarchical coding.

[0038] like Figure 1 As shown, this embodiment discloses an Alzheimer's disease diagnostic system based on graph attention and hierarchical coding, including: The image preprocessing module is configured to: acquire magnetic resonance imaging data of the object to be analyzed, and extract the cortical thickness features corresponding to each brain region after brain region division; In this embodiment, magnetic resonance imaging (MRI) data of the object to be analyzed is acquired and preprocessed, including spatial registration, intensity normalization, and noise suppression. A cortical feature extraction workflow based on brain structure analysis tools is used to divide the preprocessed MRI data into brain regions and extract cortical thickness features corresponding to each predefined brain region.

[0039] First, the raw DICOM (dcm) format magnetic resonance imaging data was converted to NIfTI (nii) format using the dcm2niigui plugin in the mricron software. The converted images were then subjected to quality checks to ensure data integrity and usability. Image quality checks were performed on the preprocessed magnetic resonance images using the SPM toolbox in MATLAB software. Each generated NIfTI image was checked for severe artifacts, missing structures, or anomalous structures, and samples with obvious quality problems were discarded.

[0040] After quality screening, the brain structure images were processed using the SBM (Surface-Based Morphometry) method in the CAT12 toolbox to extract cortical thickness features. Subsequently, the cortex was divided into multiple regions based on a pre-defined anatomical brain region template.

[0041] In some implementations, the Destrieux (a2009) atlas is used to divide the cerebral cortex into 152 regions of interest (ROIs). Among these, the "lMedial_wall" and "rMedial_wall" regions are not included in the cortical thickness quantification process as their values ​​are missing and are therefore removed in subsequent analysis, resulting in 150 valid brain region features. Each ROI corresponds to a brain region feature unit, used for subsequent brain region-level feature modeling and analysis.

[0042] The cortical thickness features corresponding to each brain region are input into the diagnostic model for processing to obtain disease diagnosis results. For example... Figure 2 The diagnostic model includes a graph attention network, a local encoder, a global encoder, and a fully connected classification layer.

[0043] The brain region feature sequence construction module is configured to: construct brain region feature sequences and adjacency matrices based on the cortical thickness features corresponding to each brain region and the anatomical adjacency relationships between brain regions; encode the brain region feature sequences based on a graph attention network, and introduce a graph attention mechanism in conjunction with the adjacency matrix to obtain updated brain region feature sequences. In this embodiment, the object to be analyzed is divided into: (e.g., 150) predefined brain regions, defining the first The average cortical thickness of each brain region is The constructed brain region feature sequence is represented as .

[0044] Constructing brain region maps , where the set of nodes Represents each predefined brain region and each node. Corresponding to a brain region Its node attributes are determined by the cortical thickness characteristics of that brain region. Representation. The set of edges in a graph structure. Used to describe the anatomical adjacency relationships between brain regions. When two brain regions share a boundary or are directly adjacent on the cortical surface, a connecting edge is considered to exist between them. For ease of subsequent calculations, the above connection relationships are represented as a weighted adjacency matrix. :

[0045] The matrix elements are defined as follows:

[0046] in, Indicates the first The node and the first Do the nodes have anatomical adjacency relationships? Indicates the first 1 node Indicates the first Each node.

[0047] To improve the stability of subsequent graph neural network computations, the adjacency matrix is ​​normalized. First, the degree matrix is ​​constructed. It is a diagonal matrix, and the diagonal elements are defined as follows:

[0048] off-diagonal elements Then the normalized adjacency matrix is ​​obtained. This matrix is ​​used to control the propagation of information about brain region features across the topology.

[0049] After obtaining the brain region feature matrix and topology matrix, a graph attention network is used to encode the brain region features. First, the raw brain region features are mapped to a high-dimensional space:

[0050] in, This represents the mapping of the original cortical thickness in a high-dimensional space, i.e., the learnable latent representation. Represents the trainable parameter matrix. This indicates the hidden feature dimension.

[0051] This mapping allows the original one-dimensional brain region thickness information to be transformed into a high-dimensional representation more suitable for deep network modeling. It should be noted that at this point... Each row still corresponds to a brain region node, but its feature form changes from the original thickness value to a learnable latent space representation.

[0052] To model the differential dependencies between brain regions, this invention further introduces a graph attention mechanism. For any central brain region... Adjacent brain regions The attention weights are calculated based on their high-dimensional representations:

[0053] in, Indicates brain regions With brain regions Attention weights , All of these are trainable parameters. For transpose, Indicates the first Initial hidden representations of brain regions. Indicates the first Initial hidden representations of brain regions.

[0054] Subsequently, for the central node Attention weights of all adjacent nodes Perform normalization to obtain the normalized adjacent nodes. Pointing to the central node Attention weights:

[0055] in, Represents the normalized brain regions With brain regions Attention weights Represents nodes The set of adjacent nodes, Indicates brain regions With brain regions Attention weights. Weights This indicates that the brain region is being updated. When representing, brain regions The proportion of information contributed can be adaptively weighted according to the importance of the relationship between brain regions.

[0056] After obtaining the normalized attention weights, information aggregation under topological constraints is performed on the brain region features to obtain the updated brain region feature representation:

[0057] in, Indicates the first Hidden representations of brain regions.

[0058] Through the aforementioned feature update process, the final representation of each brain region node not only includes the structural information of the brain region itself but also incorporates the structural context of adjacent brain regions. This enables a more accurate representation of the spatial variation patterns of neurodegenerative diseases. Compared to directly using the original brain region thickness vector, this attention-based encoding method based on brain region topology significantly enhances the model's ability to represent structural dependencies between brain regions, providing more discriminative input features for subsequent hierarchical fusion and classification prediction.

[0059] The brain region association modeling module is configured to: divide all brain regions into multiple brain region groups, obtain corresponding intra-group feature representations, input the intra-group feature representations of each brain region group into the corresponding local encoder for modeling, and obtain group-level representations; input the group-level representations into the global encoder for cross-group modeling, and obtain global brain structure representations. In this embodiment, considering that the brain is not a simple, disordered collection of brain regions, but rather has distinct anatomical divisions and functional hierarchies, this invention further constructs a hierarchical attention fusion structure based on the brain region features after graph attention encoding. Specifically, according to the anatomical affiliation or neural functional division of brain regions, all brain regions are divided into several brain region groups, such as the frontal lobe group, temporal lobe group, parietal lobe group, occipital lobe group, insula group, and limbic system-related region group, etc. Let the first... Each brain region group contains a set of brain regions. For brain region features within each functional group, self-attention modeling is performed using a local Transformer structure to learn higher-order dependencies between different brain regions within the same anatomical functional group. If the in-group input features are represented as... Then the local modeling output can be expressed as:

[0060] in, Indicates the first Local features of all brain regions within a group of brain regions. This indicates a local Transformer encoder.

[0061] The local Transformer encoder includes a multi-head self-attention module, residual connections, a layer normalization module, and a feedforward network module, which are used to enhance the feature modeling ability of intragroup brain region relationships.

[0062] After obtaining the local feature sets of all brain regions within each brain region group, in order to achieve information interaction between different brain region groups, this invention performs pooling aggregation on the local output of each brain region group to obtain a group-level representation:

[0063] in, Indicates group level representation, Indicates the first The number of brain regions contained in a group of brain regions. Indicates the first The first brain region group Local feature representations of brain regions. In this way, brain region-level features can be further compressed into group-level semantic representations, thereby reducing the complexity of global modeling while preserving important structural information within the group.

[0064] Subsequently, all group-level representations are combined in a preset order, namely the standard order commonly used in clinical and neuroscience research: frontal lobe group, parietal lobe group, temporal lobe group, occipital lobe group, insula group, and limbic system. This is then input into a global Transformer encoder for cross-group modeling to learn the interactions between different anatomical functional groups, resulting in a global brain structure representation.

[0065] in, This represents a global representation of brain structure. This represents the global Transformer encoder, which includes a multi-head self-attention module, residual connections, a layer normalization module, and a feedforward network module. This represents the group-level representation of all brain regions. This indicates the number of brain regions.

[0066] By employing a two-level modeling approach—local and global—this invention can simultaneously capture fine-grained brain region relationships and coarse-grained brain region group relationships, thereby enhancing the model's ability to understand overall brain structural patterns.

[0067] The classification prediction module is configured to input the global brain structure representation into the classification layer for prediction and obtain disease prediction results.

[0068] In this embodiment, a global brain structure representation is obtained. Subsequently, this invention achieves disease category prediction through a fully connected classification layer. Let the classification weight matrix be... The bias term is The model output is:

[0069] in, This represents the predicted probability distribution of a sample belonging to different categories. In practical applications, the classification categories can be set to three categories: Alzheimer's disease, mild cognitive impairment, and normal controls, or it can be set to a binary classification task according to actual needs.

[0070] During diagnostic model training, the true labels of the samples are recorded as... The cross-entropy loss function is used as the basic classification loss:

[0071] in, Represents cross-entropy loss, Indicates the number of categories. The one-hot encoding representing the true label of a sample, when the sample belongs to the category. hour ,otherwise By minimizing this loss, diagnostic models can be driven to learn more discriminative representations of brain structural features.

[0072] To further enhance the structural consistency of brain region features learned by the diagnostic model, this invention introduces a graph structure smoothing constraint term in addition to the classification loss. Since anatomically adjacent brain regions typically exhibit a certain continuity in structural changes, it is desirable that the latent representations of brain region pairs connected in the adjacency matrix do not show unreasonable and drastic fluctuations. Based on this, the graph smoothing loss is defined. for:

[0073] in, Indicates the first Hidden representations of brain regions. This loss can constrain the distance between adjacent brain regions in the feature space, making the representations learned by the diagnostic model more consistent with the continuous changes in brain structure.

[0074] In addition, to reduce the problem of representation instability between different brain region groups, group-level consistency constraints can be introduced. :

[0075] in, This represents the average characteristic of all group-level representations.

[0076] Finally, this invention employs a joint loss function to perform end-to-end training of the diagnostic model:

[0077] in, Denotes the joint loss function. , This is a balancing coefficient used to control the impact of different loss terms on the training process. In this embodiment, it is set to 0.1, without any specific limitation.

[0078] In this embodiment, after completing the diagnostic model training and obtaining the classification results, in order to analyze the contribution of different brain regions to the disease prediction results, the basic contribution value corresponding to each brain region feature is first calculated. Let the first brain region be the brain region with the highest contribution value. The original contribution value of each brain region is The SHAP method, based on Shapley value theory, is used to interpret and analyze the output of the Transformer network.

[0079] However, neurodegenerative diseases typically manifest as structural changes that gradually spread across multiple adjacent brain regions; therefore, considering the contribution of only a single brain region is insufficient to reflect the structural dependencies between brain regions. To address this, this invention incorporates brain region topological information. For the original contribution value Structural propagation corrections are performed to construct a structural perception contribution index. This provides a numerical value indicating the extent to which each brain region contributes to the classification results. This is helpful for doctors in making diagnoses. A larger value indicates a greater impact on the classification.

[0080] The above process can be expressed by the following formula.

[0081] in, This represents the propagation coefficient, used to control the degree of influence of the contribution values ​​of adjacent brain regions on the contribution of the current brain region. In this embodiment, it is set to 0.05, without any specific limitation.

[0082] This approach allows for the introduction of brain region topological information while maintaining the interpretability of the original contribution values, making the resulting contribution index more consistent with the biological patterns of neurodegenerative diseases spreading between brain regions.

[0083] This invention utilizes the self-attention mechanism of Transformer networks to achieve explicit modeling of dependencies between different brain regions, thereby improving the accuracy of disease classification. The SHAP theory is applied to the brain region representation after attention encoding, enabling brain region importance analysis to reflect synergistic effects rather than isolated feature contributions. While ensuring model performance, this invention improves the interpretability of model decision results and enhances its clinical application value.

[0084] In summary, the diagnostic system provided by this invention has the following beneficial effects: This invention introduces a graph attention network, which adaptively learns the influence weights of different neighboring brain regions on the target brain region, achieving effective information aggregation of cortical thickness features in the brain region topology. Compared to traditional methods that ignore brain region relationships or use fixed weights, this invention can more accurately capture the spatial propagation patterns of Alzheimer's disease pathological changes, thereby obtaining more discriminative brain region feature representations and effectively improving the accuracy of disease classification.

[0085] This invention designs a two-level coding architecture, in which the local encoder performs fine-grained modeling of the brain region relationships within each brain region group, while the global encoder performs cross-group modeling of the group-level representations of different brain regions. This hierarchical processing method abstracts layer by layer from fine-grained to coarse-grained, preserving the detailed changes in local brain regions while integrating the structural context at the whole-brain scale, resulting in a more comprehensive and biologically plausible global brain structure representation.

[0086] This invention innovatively introduces smoothing constraints based on brain region topology and group-level consistency constraints, building upon traditional classification loss. These constraints ensure that the feature representations learned by the model conform to the continuity of changes in the cerebral cortex structure and the consistency within functional groups, thereby significantly improving the model's training stability, resistance to noise, and generalization performance across different datasets.

[0087] This invention proposes a structural perception contribution index, which uses the topological adjacency relationship of brain regions to make propagation corrections based on the baseline contribution value. This index not only reflects the importance of a single brain region, but also reflects the combined influence of its neighboring brain regions, thus more accurately characterizing the diffusion effect of Alzheimer's disease pathology between adjacent brain regions.

[0088] The experiment is described below: The dataset includes 69 patients with Alzheimer's disease (AD), 50 patients with mild cognitive impairment (MCI), and 82 healthy elderly volunteers (NC), who are divided into training and validation groups in a 7:3 ratio.

[0089] Evaluation indicators: Table 1 Confusion Matrix

[0090] True positives (TP): The number of samples correctly classified as positive, i.e., the number of samples that are actually positive and are classified as positive by the model; False positives (FP): The number of samples incorrectly classified as positive, i.e., the number of samples that are actually negative but are classified as positive by the model; False negatives (FN): The number of samples incorrectly classified as negative, i.e., the number of samples that are actually positive but are classified as negative by the model; True negatives (TN): The number of samples correctly classified as negative, i.e., the number of samples that are actually negative and are classified as negative by the model.

[0091] The reserve ratio is expressed as:

[0092] in, This represents the percentage of the total samples where the model correctly predicted the results, i.e., the sum of the sample trees where the actual result was positive but the prediction was positive, and the sample trees where the actual result was negative but the prediction was negative. It is an important metric for evaluating the model.

[0093] Sensitivity is expressed as:

[0094] in, Also known as sensitivity, true positive rate, or recall, it represents the proportion of all positive cases that are correctly classified. It measures the classifier's ability to identify positive cases, and in medicine, it can be expressed as the ability to identify patients among real patients. The false negative rate is negatively correlated with the false negative rate. False negative rate = 1 - sensitivity.

[0095] Specificity is expressed as:

[0096] in, Also known as the true negative rate, it represents the proportion of all negative cases that are correctly classified. It measures the classifier's ability to identify negative cases. In medicine, it refers to the proportion of healthy people who are diagnosed as healthy. It is negatively correlated with the false diagnosis rate. False diagnosis rate = 1 - specificity.

[0097] The F1 index is expressed as:

[0098]

[0099]

[0100] in, Indicates accuracy. The F1 score represents recall. It can be seen as a weighted average of model precision and recall, comprehensively considering both metrics, and is often used to evaluate model precision. The F1 score ranges from zero to one; a value closer to 1 indicates a better model.

[0101] The area under the ROC curve refers to the area enclosed by the ROC curve and the horizontal axis. The ROC curve is calculated using the parameters of the confusion matrix. The threshold for a binary classification model can be chosen to be high or low. The threshold essentially refers to the probability (a value between 0.0 and 1.0) that the model outputs for each sample; observations with scores equal to or higher than this threshold produce a positive output. Choosing different thresholds will result in different false positive rates (FPR) and true positive rates (TPR). Plotting the FPR for each threshold of the same model on the horizontal axis and the TPR on the vertical axis in ROC space creates the ROC curve for that specific model. Generally, a classifier with a higher AUC value has a higher accuracy.

[0102] True positive rate (TPR):

[0103] in, This indicates sensitivity, representing the proportion of positive examples that are correctly identified.

[0104] False positive rate (FPR):

[0105] in, The probability of being classified as a positive example but not a true negative example, which is the probability of being classified as a positive example among true negative examples.

[0106] The proposed network was tested on four classification tasks: a three-class classification task (AD, MCI, NC) and three binary classification tasks (MCI and NC, AD and NC, AD and MCI). The experimental results are shown in Table 2 below, with the confusion matrix for each task listed in the table. Figure 3 , Figure 4 , Figure 5 , Figure 6 It is displayed in the middle.

[0107] Table 2 Classification results of Transformer structures

[0108] The data in the table shows that this algorithm has the lowest accuracy and sensitivity in the three-class classification task. It achieves the highest accuracy in the AD and MCI classification tasks, but its sensitivity and specificity are not the highest, resulting in relatively average performance.

[0109] For the MCI and NC classification tasks, although the sensitivity was the highest, the specificity was the lowest, indicating a low false negative rate but a high false negative rate. For the AD and NC classification and diagnosis, the accuracy was the lowest, but the specificity was the highest, indicating a relatively low false negative rate. Furthermore, the F1 score was around 0.9 in each classification task, indicating high model quality and a good balance between accuracy and sensitivity. The area under the curve was above 0.84, indicating high realism of the detection method.

[0110] Figure 3 The confusion matrix is ​​more intuitive than the accuracy and sensitivity data in Table 2. Figure 3 The results show that the classification of AD is relatively good, with most AD cases correctly identified. However, two cases were missed and misdiagnosed as MCI and NC. The classification of MCI and NC is not as good as AD, with NC having a higher error rate. Specifically, MCI was mainly identified as NC, indicating a missed diagnosis. NC was mainly identified as AD, indicating a misdiagnosis. The three-class classification performance is indeed not very good, and there is still significant room for improvement.

[0111] from Figure 4 As can be seen, diagnosing MCI is quite difficult. MCI is mild cognitive impairment, which is in the early stage of the disease. The pathological changes may not be obvious, and it is difficult to distinguish it from normal people. However, this is an important step in identifying the disease in its early stages and is of great significance for preventing the progression of the disease. Therefore, the model still needs to be further improved in this classification task.

[0112] from Figure 5 As can be seen, AD and NC are classified relatively well because they are located at opposite ends of the disease progression, with significant differences in cortical thickness, making them easier to distinguish. Although 5 normal individuals were misclassified as patients, this was corrected by doctors in subsequent disease diagnoses. The low rate of missed diagnoses in this classification task indicates that the model can effectively identify patients.

[0113] from Figure 6 As can be seen, AD and MCI are classified better than MCI and NC. This may be because most MCI patients in the dataset do not have obvious symptoms and have not reached the late stage of MCI, so the difference between them and AD is large, but the difference between them and NC is small.

[0114] For the comparison algorithms, two deep learning algorithms were selected: Convolutional Neural Networks (CNN) and a variant of recurrent neural networks, Long Short-Term Memory (LSTM). Three machine learning algorithms were also selected: Support Vector Machines (SVM), Naïve Bayes (NB), and Decision Trees (DT). Experiments were conducted on three-class classification tasks and three two-class classification tasks. The experimental results are shown in Tables 3, 4, 5, and 6.

[0115] Table 3 Comparison of Algorithms in AD, MCI, and NC Classification Tasks

[0116] Table 3 shows that the algorithm of this invention achieves better accuracy than other comparative algorithms in the three-class classification task, and is similar in accuracy to the LSTM algorithm. However, it surpasses the LSTM algorithm in terms of sensitivity and specificity, indicating that the model has a strong ability to detect patients and a low misdiagnosis rate. In this classification task, the specificity of all algorithms is higher than their sensitivity, indicating that the model has a low misdiagnosis rate but some missed diagnoses.

[0117] Table 4 Comparison of algorithms in MCI and NC classification tasks

[0118] Table 4 shows that the algorithm of this invention outperforms other comparative algorithms in terms of accuracy, and is similar to the Naive Bayes algorithm in accuracy, but surpasses it in sensitivity, indicating a stronger ability to identify patients. While the specificity of the model in this invention is not the highest, models with higher specificity have lower sensitivity. Furthermore, in this classification task, the accuracy of both deep learning algorithms is not as high as that of the machine learning algorithms.

[0119] Table 5 Comparison of Algorithms in AD and NC Classification Tasks

[0120] Table 5 shows that both deep learning networks improved accuracy compared to machine learning algorithms. However, the algorithm proposed in this invention achieves higher accuracy, exceeding the others by more than 10%. Furthermore, in this classification task, the five compared algorithms all showed higher sensitivity than specificity, indicating a strong ability to detect patients, but with a certain false positive rate. In contrast, the model in this invention exhibits higher specificity than sensitivity, indicating a lower false positive rate, but with a certain false negative rate, which requires improvement.

[0121] Table 6 Comparison of Algorithms in AD and MCI Classification Tasks

[0122] Table 6 shows that both deep learning networks improve accuracy compared to machine learning algorithms. However, the algorithm proposed in this paper achieves higher accuracy. Furthermore, all algorithms demonstrate higher sensitivity than specificity in this classification task, indicating a strong ability to identify patients, but with a certain rate of false diagnoses. The model in this invention exhibits the highest specificity, thus reducing the false diagnoses rate.

[0123] The SHAP method is used to identify brain regions that contribute significantly to the classification task in the Transformer algorithm after training for each classification task. These regions are then ranked, visualized, and used to locate lesion brain regions. Figure 7 , Figure 8 , Figure 9 , Figure 10 As shown.

[0124] Based on the correspondence between brain region numbers and templates, the following are the names of the brain regions: 77 corresponds to the right middle temporal gyrus (rG_temporal_middle); 135 corresponds to the right pericallosal sulcus (rS_pericallosal); 76 corresponds to the left middle temporal gyrus (lG_temporal_inf); 70 corresponds to the polar plane of the left superior temporal gyrus (lG_temp_sup-Plan_polar); 88 corresponds to the left temporal pole (lPole_temporal); 36 corresponds to the left short insular gyrus (lG_insular_short); 37 corresponds to the right short insular gyrus (rG_insular_short); 122 corresponds to the left lateral occipitotemporal sulcus (lS_oc-temp_lat); 81 corresponds to the anterior vertical branch of the right lateral sulcus (or fissure) (rLat_Fis-ant-Vertical); 141 corresponds to the upper part of the right precentral sulcus (rS_precentral-sup-part).

[0125] The temporal and frontal lobes are the main sites affected after the onset of Alzheimer's disease. Figure 7 The brain regions that contribute significantly to the classification process are mostly located in the temporal lobe. The length of each color in the diagram represents its proportion in different categories. For example, region 77 has the largest proportion in category 0, i.e., AD.

[0126] Analysis of variance (ANOVA) was performed on the brain regions corresponding to the first two groups using SPSS software. ANOVA is a method for comparing multiple groups, allowing for the comparison of differences between three or more groups. The results are shown in Table 7. Analysis of the cortical thickness of the brain region corresponding to 77 showed a statistically significant P-value (P < 0.05). Similarly, analysis of the cortical thickness of the brain region corresponding to 135 also showed a statistically significant P-value (P < 0.05).

[0127] Table 7. Results of ANOVA on brain regions for AD, MCI, and NC tri-classification.

[0128] Based on the correspondence between brain region numbers and templates, the following are the names of the brain regions: 70 corresponds to the polar plane of the left superior temporal gyrus in 'lG_temp_sup-Plan_polar'; 65 corresponds to the right inferior callosal gyrus in 'rG_subcallosal'; 64 corresponds to the left inferior callosal gyrus in 'lG_subcallosal'; 135 corresponds to the right pericallosal groove in 'rS_pericallosal'; 37 corresponds to the right short insular gyrus in 'rG_insular_short'; 63 corresponds to the right straight gyrus in 'rG_rectus'; 57 corresponds to the right postcentral gyrus in 'rG_postcentral'; 21 corresponds to the right posteroventral part of the cingulum gyrus in 'rG_cingul-Post-ventral' (vPCC, isthmus of the cingulum gyrus); 59 corresponds to the right precentral gyrus in 'rG_precentral'; 71 corresponds to the polar plane of the right superior temporal gyrus in 'rG_temp_sup-Plan_polar'.

[0129] Using SPSS software, group statistics were analyzed for the first two brain regions, and the results are shown in Table 8. Analysis of the cortical thickness of the brain region corresponding to 70 showed that the average cortical thickness for the MCI category was 3.237652264, and the average for the NC category was 3.250458518. Similarly, analysis of the cortical thickness of the brain region corresponding to 65 showed that the average cortical thickness for the MCI category was 2.925281356, and the average for the NC category was 2.740888546.

[0130] Table 8 Group statistics of brain regions classified as MCI and NC

[0131] The mean equation t-test was performed on the two brain regions, and the results are shown in Table 3-10. In the Roi70 test, P>0.05, indicating no statistical significance. This suggests that the differences between MCI and NC are small, making classification difficult. In the Roi65 test, P<0.05, indicating statistical significance.

[0132] Table 9. Results of the T-test for the mean equation of brain regions in MCI and NC classifications.

[0133] Based on the correspondence between brain region numbers and templates, the following are the names of the brain regions: 34 corresponds to the left long insular gyrus and central insular sulcus of 'lG_Ins_lg_and_S_cent_ins'; 92 corresponds to the left central sulcus of 'lS_central'; 64 corresponds to the left subcallosal gyrus of 'lG_subcallosal'; 67 corresponds to the right anterior transverse temporal gyrus of 'rG_temp_sup-G_T_transv'; 100 corresponds to the upper end of the circular sulcus of the left insula of 'lS_circular_insula_sup'; 73 corresponds to the temporal plane of the right superior temporal gyrus of 'rG_temp_sup-Plan_tempo'; 126 corresponds to the left lateral orbital sulcus of 'lS_orbital_lateral'; 77 corresponds to the right middle temporal gyrus of 'rG_temporal_middle'; 116 corresponds to the left middle occipital sulcus and lunate sulcus of 'lS_oc_middle_and_Lunatus'; 135 corresponds to the right pericallosal sulcus of 'rS_pericallosal'.

[0134] Figure 10 In the AD and NC classification tasks, six brain regions were located on the left side, which is consistent with the findings by Roe JM et al. that the left brain atrophies faster than the right brain in Alzheimer's disease.

[0135] Using SPSS software, group statistics were analyzed for the first two brain regions, and the results are shown in Table 3-11. Analysis of the cortical thickness of the brain region corresponding to 34 showed that the average cortical thickness for the AD category was 3.052133409, and the average for the NC category was 3.223341279. Analysis of the cortical thickness of the brain region corresponding to 92 showed that the average cortical thickness for the AD category was 1.560209375, and the average for the NC category was 1.669837921. Both brain regions showed a decrease in cortical thickness in the AD category.

[0136] Table 10 Group statistics of brain regions classified as AD and NC

[0137] The mean equation t-test was performed on the two brain regions, and the results are shown in Table 11. In the ROI 92 test, P < 0.05, indicating statistical significance. In the ROI 65 test, P < 0.05, indicating statistical significance.

[0138] Table 11 Results of the T-test for the mean equation of brain regions in AD and NC classifications

[0139] Based on the correspondence between brain region numbers and templates, the following are the names of the brain regions: 44 corresponds to the left lingual gyrus (the language portion of the medial occipitotemporal gyrus) of 'lG_oc-temp_med-Lingual'; 45 corresponds to the right lingual gyrus (the language portion of the medial occipitotemporal gyrus) of 'rG_oc-temp_med-Lingual'; 36 corresponds to the left short insular gyrus of 'lG_insular_short'; 37 corresponds to the right short insular gyrus of 'rG_insular_short'; 97 corresponds to the anterior segment of the circular sulcus of the right insula of 'rS_circular_insula_ant'; 102 corresponds to the left anterior transverse accessory sulcus of 'lS_collat_transv_ant'; 38 corresponds to the left middle occipital gyrus of 'lG_occipital_middle'; 109 corresponds to the right middle frontal sulcus of 'rS_front_middle'; 52 corresponds to the left superior marginal gyrus of 'lG_pariet_inf-Supramar'; 101 corresponds to the upper end of the circular sulcus of the right insula of 'rS_circular_insula_sup'.

[0140] Using SPSS software, group statistics were analyzed for the first two brain regions, and the results are shown in Table 3-13. Analysis of the cortical thickness of the brain region corresponding to 44 showed that the average cortical thickness in the AD category was 1.670581283, and the average in the MCI category was 1.945356214. Similarly, analysis of the cortical thickness of the brain region corresponding to 45 showed that the average cortical thickness in the AD category was 1.839742780, and the average in the MCI category was 1.991225404.

[0141] Table 12 Group statistics of brain regions classified as AD and MCI

[0142] The mean equation t-test was performed on the two brain regions, and the results are shown in Table 3-12. In the Roi 44 test, P < 0.05 was considered statistically significant. In the Roi 45 test, P < 0.05 was also considered statistically significant.

[0143] Table 13 Results of the T-test for the mean equation of brain regions classified as AD and MCI

[0144] Example 2 The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the contents of the system of Embodiment 1.

[0145] Example 4 The purpose of this embodiment is to provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, performs the contents of the system of Embodiment 1.

[0146] The contents involved in the apparatus of Embodiments 3 and 4 above correspond to those of System Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to execute any instance of the present invention.

[0147] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0148] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0149] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. An Alzheimer's disease diagnostic system based on graph attention and hierarchical coding, characterized in that, include: The image preprocessing module is configured to: acquire magnetic resonance imaging data of the object to be analyzed, and extract the cortical thickness features corresponding to each brain region after brain region division; The brain region feature sequence construction module is configured to: construct brain region feature sequences and adjacency matrices based on the cortical thickness features corresponding to each brain region and the anatomical adjacency relationships between brain regions; encode the brain region feature sequences based on a graph attention network, and introduce a graph attention mechanism in conjunction with the adjacency matrix to obtain updated brain region feature sequences. The brain region association modeling module is configured to: divide all brain regions into multiple brain region groups, obtain corresponding intra-group feature representations, input the intra-group feature representations of each brain region group into the corresponding local encoder for modeling, and obtain group-level representations; input the group-level representations into the global encoder for cross-group modeling, and obtain global brain structure representations. The classification prediction module is configured to input the global brain structure representation into the classification layer for prediction and obtain disease prediction results.

2. The Alzheimer's disease diagnostic system based on graph attention and hierarchical coding as described in claim 1, characterized in that, A degree matrix is ​​constructed based on the adjacency matrix, and the adjacency matrix is ​​normalized based on the degree matrix to obtain a normalized adjacency matrix.

3. The Alzheimer's disease diagnostic system based on graph attention and hierarchical coding as described in claim 2, characterized in that, The updated brain region feature sequence was obtained, specifically including: The brain region feature sequences are mapped to a high-dimensional space to obtain the initial hidden representation of the brain region; By combining the adjacency matrix, the attention weights of each brain region and its neighboring brain regions are calculated based on the initial hidden representation of the brain region. After normalization, the normalized attention weights are obtained. Based on the normalized attention weights, information aggregation under topological constraints is performed on the brain region feature sequences to obtain the updated brain region feature sequences.

4. The Alzheimer's disease diagnostic system based on graph attention and hierarchical coding as described in claim 1, characterized in that, Based on the anatomical affiliation of brain regions, all brain regions are divided into multiple brain region groups.

5. The Alzheimer's disease diagnostic system based on graph attention and hierarchical coding as described in claim 1, characterized in that, The group-level representation is obtained by: inputting the feature representations within each brain region group into the corresponding local encoder for modeling, thereby obtaining the local feature representations of all brain regions within each brain region group; and pooling and aggregating the local feature representations of each brain region group to obtain the group-level representation.

6. The Alzheimer's disease diagnostic system based on graph attention and hierarchical coding as described in claim 1, characterized in that, The global brain structure representation is obtained, which specifically includes: combining all group-level representations in a preset order and inputting them into the global encoder for cross-group modeling.

7. The Alzheimer's disease diagnostic system based on graph attention and hierarchical coding as described in claim 1, characterized in that, The model is trained end-to-end using a joint loss function, as follows: in, Denotes the joint loss function. , Represents the balance coefficient. Represents cross-entropy loss, This represents the graph smoothing loss. This represents the group-level consistency constraint loss.

8. The Alzheimer's disease diagnostic system based on graph attention and hierarchical coding as described in claim 7, characterized in that, The group-level consistency constraint loss is expressed as: in, This represents the group-level consistency constraint loss. Indicates the first Group-level representation of brain regions This represents the average characteristic of all group-level representations.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the content of the Alzheimer's disease diagnostic system based on graph attention and hierarchical coding as described in any one of claims 1-8.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the content of the Alzheimer's disease diagnostic system based on graph attention and hierarchical coding as described in any one of claims 1-8.

Images