Alzheimer's disease diagnosis system based on interpretable graph neural network

Abstract

The invention relates to an interpretable graph neural network-based Alzheimer's disease diagnosis system, which belongs to the technical field of medical internet of things and comprises a preprocessing module, a local graph construction module, an important feature node selection module, a node selection process display module, a global graph construction module, an Alzheimer's disease diagnosis module and a global attribution interpretation module. Extracting important feature nodes in the neural image data based on a self-attention pooling mechanism, and explaining a local feature selection process in detail; the selected important node information is integrated with non-imaging data, the Alzheimer's disease diagnosis module is utilized to predict three categories of Alzheimer's disease, and finally the global attribution interpretation module is utilized to perform attribution interpretation on global and single sample features to verify the validity of feature selection. According to the method, the diagnosis accuracy is improved, the good interpretability is achieved, and a more reliable and effective diagnosis auxiliary tool can be provided for clinicians.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical Internet of Things and relates to an Alzheimer's disease diagnosis system based on an interpretable graph neural network. Background Art

[0002] Alzheimer's disease (AD) is an irreversible neurodegenerative disease, and its main clinical manifestations are the continuous decline of memory, cognitive function and physical ability. According to the patient's physical condition and cognitive level, the progression of AD can be divided into three stages: normal control, mild cognitive impairment and AD. In clinical diagnosis, the early symptoms of AD and MCI are very hidden, making it extremely difficult to identify the disease early. In particular, MCI is a transitional state between normal aging and Alzheimer's disease, with a milder degree of cognitive decline, which is difficult to distinguish from normal aging. Traditionally, the diagnosis of AD relies on the doctor's clinical experience, cognitive function assessment, and neuroimaging examinations. However, due to large individual differences, conventional diagnostic methods are difficult to achieve accurate diagnosis in the early stages of the disease.

[0003] In recent years, many researchers have relied on single unimodal data to predict AD, such as magnetic resonance imaging (MRI), structural magnetic resonance imaging (sMRI), and positron emission tomography (PET). Although this method has achieved good results in some cases, unimodal data often cannot fully capture the complex case process of AD, which limits the prediction accuracy of the model. Therefore, multimodal data has gradually become the focus of AD research, which can provide more comprehensive case information. The integration of these multimodal data can reveal the complex pathological mechanisms of AD from different angles, thereby greatly improving the accuracy of prediction. At the same time, although multimodality can improve the accuracy of prediction, due to the complex potential relationships between different types of data, simple feature splicing or fusion methods based on homogeneous features are difficult to capture the deep connection between data. Therefore, how to effectively fuse multimodal data and build a highly adaptable model has become the key to the early diagnosis of Alzheimer's disease.

[0004] In addition, although deep learning models can achieve excellent performance in various tasks through a large number of parameters and complex hierarchical structures, these models are often regarded as "black boxes" and it is difficult to interpret their internal decision-making processes. Especially in clinical scenarios involving patient health, doctors need to understand the decision-making basis of the model in order to verify and explain the model's prediction results. Therefore, in the early prediction of AD, it is particularly important to enhance the interpretability of the model. Summary of the invention

[0005] In view of this, the purpose of the present invention is to provide a novel multimodal data fusion prediction framework based on local-global interpretable graph neural network to meet the current challenges of early diagnosis of AD and solve the problems of early diagnosis and interpretability of AD.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] An Alzheimer's disease diagnosis system based on an interpretable graph neural network, comprising:

[0008] Preprocessing module: used to preprocess the original neuroimaging data to obtain standard T1 images;

[0009] Local map construction module: used to segment the standard T1 image into multiple regions of interest to form a local map;

[0010] Important feature node selection module: It is used to take the local graph as input and select important feature nodes through a self-attention-based pooling mechanism;

[0011] Node selection process display module: Use the top-k selection strategy in the pooling mechanism combined with the heat map to explain the important feature node selection process and results;

[0012] Global graph construction module: used to construct data samples by combining selected important node features and non-imaging features, and then randomly select known samples and unknown samples from the data set to form a global graph;

[0013] Alzheimer's disease diagnosis module: used to use the global graph as the input of the GNN layer, adaptively learn the relationship between nodes, predict Alzheimer's disease, and obtain the diagnosis result of Alzheimer's disease;

[0014] Global Attribution Explanation Module: Visualize the impact of features on results through interpretable methods to understand diagnostic decision-making behavior.

[0015] Furthermore, the preprocessing module preprocesses the original neuroimaging data, specifically including skull removal, cerebellum resection, image segmentation, standardization, and quality monitoring, so as to obtain a standard T1 image.

[0016] Furthermore, the local graph construction module first uses the Harvard-Oxford atlas to segment the preprocessed standard T1 image into multiple regions of interest (ROIs), each ROI corresponds to a node, to construct a local graph, and simultaneously obtains the feature matrix of each node and the adjacency matrix of the local graph.

[0017] Furthermore, the important feature node selection module performs a pooling operation according to the topological structure of the importance of the nodes, that is, selects the nodes using a top-k strategy and discards unimportant nodes.

[0018] Furthermore, the node selection process display module combines the top-k strategy with the heat map, and uses the heat map to present the nodes after the scores are assigned and the nodes after the selection, so as to display the node selection process.

[0019] Furthermore, the global graph construction module fuses the new imaging node data with the non-imaging data to form a sample set of the global graph, randomly selects α known samples and β unknown samples from the sample set as the input S of the graph neural network module, and uses the input S to establish relationships on the global graph.

[0020] Furthermore, the Alzheimer's disease diagnosis module takes the global graph as input and calculates the edge probability matrix and the edge weight matrix, specifically including:

[0021] Calculate the edge probability matrix E using prior knowledge and risk factors:

[0022]

[0023] Where k represents the kth risk factor feature, K is the number of risk factor features, and e i,j is the i-th row and j-th column of the edge probability matrix E, representing the edge weights between nodes;

[0024] Then, the CNN model is used to calculate the similarity of different nodes in the node set, obtain the absolute difference between each feature of the two nodes, and learn the edge weights from the absolute difference to obtain the edge weight matrix

[0025]

[0026] in, Represents the edge weight matrix W of the l-th layer network l The weights between nodes in ;

[0027] This results in a self-metric adjacency matrix with probability constraints:

[0028]

[0029] Represents element-wise multiplication;

[0030] Use the SOFTMAX layer to normalize the matrix and add it to the adjacency operator family after normalization. In order to update the node; D represents the diagonal matrix, which is used to preserve the characteristics of the node itself when updating the node;

[0031] Use each operator in the operator family to update the node and accumulate the update results G n (V (l) ), and then add the accumulated result to the input node V of this layer (l) Connect and retain the characteristics of the input node while obtaining the output node V of the GNN layer (l+1) =[V (l) , G n (V (l) )];

[0032] Finally, after the feature dimension is changed in the last layer, the output of the unknown node is directly input into the SOFTMAX layer to normalize the output; the normalized result is the prediction of the unknown node, thus completing the diagnosis prediction of Alzheimer's disease.

[0033] Furthermore, the global attribution explanation module performs global attribution explanation on the trained model through a model-agnostic local proxy interpretable fusion method, that is, SHAP is used to explain the impact of global features on the model output, and the LIME method is used to analyze the impact of each feature on different categories; thereby explaining the impact of each feature on the prediction result and understanding the diagnostic decision-making behavior of the model.

[0034] The beneficial effects of the present invention are: by introducing the Harvard-Oxford atlas to divide brain regions, and combining the self-attention mechanism to extract key feature nodes in the neuroimaging image, it can effectively integrate imaging and non-imaging data, and achieve comprehensive and accurate AD diagnosis prediction; in the local processing part, the feature selection process is explained in detail through the Top-k selection strategy and heat map, and in the global prediction part, SHAP and LIME are used to perform feature interpretation analysis, revealing the impact of features on model output, and further verifying the effectiveness of feature selection and the impact of each feature on model decision. Therefore, the present invention can not only have a high prediction accuracy, but also has good generalization ability and good interpretability.

[0035] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0037] Figure 1 This is the architecture diagram of the Alzheimer's disease diagnosis system based on interpretable graph neural network;

[0038] Figure 2 Flowchart of the Alzheimer's disease diagnosis method based on interpretable graph neural network. DETAILED DESCRIPTION

[0039] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0040] It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and thus the drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0041] In the following description, numerous details are discussed to provide a more thorough explanation of the embodiments of the present invention. However, it is obvious to those skilled in the art that the embodiments of the present invention can be implemented without these specific details. In other embodiments, well-known structures and devices are shown in the form of block diagrams rather than in detail to avoid making the embodiments of the present invention difficult to understand.

[0042] The present invention provides an Alzheimer's disease diagnosis system based on a local-global interpretable graph neural network to improve the AD prediction accuracy while increasing the interpretability.

[0043] like Figure 1-2 As shown in the figure, in order to reduce the complexity of the prediction system model execution, the important feature node selection module based on self-attention is used to classify the important nodes of the imaging image. Then, its output results are simultaneously passed to the Alzheimer's disease diagnosis module with non-imaging data, and the disease state is predicted by generating a global node graph and using the GNN model for classification. In addition, in order to make the prediction process more visual and interpretable, a local-global interpreter is used, including a node selection process display module and a global attribution explanation module. The node selection process display module is mainly responsible for visualizing the decision of important features, while the global attribution explanation module explains the entire prediction process to ensure that the decision made by the method is more reliable and convincing.

[0044] The specific steps of the system are as follows:

[0045] Preprocessing module: preprocess the original neuroimaging data to obtain standard T1 images.

[0046] Due to the particularity of early AD prediction research, the present invention is based on the ADNI database of Alzheimer's neuroimaging technology for research. 2000 samples were randomly selected from the database, including data features such as sMRI image data, age, gender, cognitive scores, etc., with the aim of using multimodal features to more accurately predict AD. Since the selected sMRI is raw data, it contains a large amount of redundant information. Therefore, before importing the data, the image data needs to be preprocessed. The statistical parameter mapping (Spm 12) tool is mainly used to remove the skull, cerebellum, image segmentation, standardization, and quality monitoring of the data to obtain standard T1 images, so as to effectively reduce the time cost of model processing data and make the results more accurate.

[0047] Local map construction module: The standard T1 image is segmented into multiple regions of interest (ROIs) to form a local map.

[0048] In order to more accurately map the imaging data to specific areas, the Harvard-Oxford atlas was used to segment the subject's brain described by the sMRI image into n regions of interest (ROIs), each ROI corresponding to a node to construct a local graph. Correspondingly, the feature matrix of each node was obtained, and the adjacency matrix of the node can be defined as 1 if two nodes are adjacent, and 0 otherwise.

[0049] Important feature node selection module: takes the local graph as input and selects important feature nodes through a self-attention based pooling mechanism.

[0050] The topological structure of the nodes is pooled according to their importance, that is, the top-k strategy is used to select nodes, and (1-k)N less important nodes are discarded, where the pooling ratio is the hyperparameter k∈(0,1], which determines the number of nodes to be retained, and N is the number of nodes. The adjacency matrix and node features are updated to obtain a new adjacency matrix and feature matrix, and the output result of the local selection module is obtained, and the output feature where x i is the feature vector of the i-th node, and || represents the connection, thereby completing the node selection and obtaining the data information of important nodes.

[0051] Node selection process display module: The top-k selection strategy in the pooling mechanism is combined with the heat map to explain the important feature node selection process and results.

[0052] The self-attention pooling mechanism used in the important node selection process can determine the important nodes by calculating the weight for each input node, that is, calculating the self-attention score for each node, and using the top-k selection strategy to select the top kN nodes. Although this mechanism and strategy itself has a certain degree of interpretability, different imaging methods can provide necessary supplementary information to improve the accuracy of disease diagnosis. Therefore, the top-k strategy is combined with the heat map, and the nodes after the scores are assigned and the nodes after the selection are presented using the heat map, so as to show the node selection process, enhance the understanding of the decision-making behavior of the top-k strategy, reduce unnecessary data analysis and calculation processes, and provide users with a more transparent and explainable decision-making process.

[0053] Global graph construction module: The selected important node features and non-imaging features form data samples, and then known samples and unknown samples are randomly selected from the data set to form a global graph.

[0054] The new imaging node data is fused with the non-imaging data (age, gender, APOE4 gene information, cognitive test score features) to form a sample a of the global graph. The number of nodes in the graph G of each sample is N = s + 3 + 1, where s is the number of important points output by the local selection module, 3 represents age, gender, APOE4 gene information, and 1 represents cognitive test score features. Then, 10 known samples and 1 unknown sample are randomly selected from the sample set as the input S of the model. c is the label, C is the number of categories (AD early involved three categories: NC, MCI, AD), and the input S is used to establish the relationship between the global graph G = {V, E, W}, that is, the node v i The initialization features are updated, and the initialized graph is set to be fully connected, and the weights of its edges are all 1.

[0055] Alzheimer's disease diagnosis module: The global graph is used as the input of the model, and AD prediction is performed through a graph neural network that can adaptively learn the relationship between nodes to obtain AD diagnosis results.

[0056] The global graph is input into the GNN layer, and the node features are updated through information transmission between nodes to obtain the labels of unknown nodes and realize the diagnosis and prediction of AD. It mainly includes two steps: calculation of the self-metric connection matrix and node update. Among them, since the input data is multimodal, that is, it contains risk factors, MRI data, label information, etc., it is necessary to calculate the edge probability matrix and the edge weight matrix separately before obtaining the self-metric connection matrix with probability constraints. First, the edge probability matrix E is calculated using prior knowledge and risk factors. Since the risk factors have a low dimension and a regular numerical distribution, they are easily affected by other features. Therefore, they can be calculated directly based on prior knowledge. The edge probability matrix calculated using risk factors is Where k represents the kth risk factor feature, K is the number of risk factor features, and e i,j is the i-th row and j-th column of the edge probability matrix E, which represents the edge weights between nodes. Then, the CNN model is used to calculate the similarity of different nodes in the node set, obtain the absolute difference between each feature of the two nodes, and learn the edge weights from the absolute difference to obtain the edge weight matrix in, Represents the edge weight matrix W of the l-th layer network l The weights between nodes in . Thus, a self-metric adjacency matrix with probability restriction can be obtained Represents element-wise multiplication, and uses the SOFTMAX layer to normalize the matrix, which is then added to the adjacency operator family. In order to update the node. D represents the diagonal matrix, which is used to preserve the characteristics of the node itself when updating the node.

[0057] Use each operator in the operator family to update the node and accumulate the update results G n (V (l) ), and then add the accumulated result to the input node V of this layer (l) Connect and retain the characteristics of the input node while obtaining the output node V of the GNN layer (l+1) =[V (l) , G n (V (l) )]. Finally, after the feature dimension is changed in the last layer, the output of the unknown node is directly input into the SOFTMAX layer to normalize the output. The normalized result is the prediction of the unknown node, thus completing the diagnosis prediction of AD.

[0058] Global attribution explanation module: Visualize the impact of features on results through interpretable methods and understand the diagnostic decision-making behavior of the model

[0059] The trained model is globally attributed and explained through a model-agnostic local proxy interpretable fusion method, that is, SHAP is used to explain the impact of global features on model output. At the same time, in order to more deeply analyze the impact of each feature on each sample, the LIME method is used to analyze the impact of each feature on different categories. By combining the two methods, the impact of each feature on the prediction results can be explained, and the diagnostic decision-making behavior of the model can be understood, so as to better assist doctors in diagnosis and prediction.

[0060] Based on explainable artificial intelligence technology and deep learning methods, the present invention aims at the accuracy and explainability of early diagnosis prediction of Alzheimer's disease, and proposes a diagnostic method of interpretable graph neural network for early Alzheimer's patients. The imaging data is locally selected and fused with non-imaging data to improve the diagnostic accuracy of the model. At the same time, in order to increase the interpretability of the method, the local selection module adopts the combination of the top-k selection strategy, which is interpretable in itself, and the heat map, so that the imaging data node selection process is transparent. At the same time, two interpretable methods are used to perform attribution analysis on the output of the model to determine the single sample features and global features that have the greatest impact on the decision results, so as to assist doctors in understanding the decision-making process.

[0061] The present invention proposes a local-global interpretable graph neural network method that integrates multimodal data to perform three-classification prediction for early AD diagnosis. By introducing the Harvard-Oxford atlas to divide brain regions and combining the self-attention mechanism to extract key feature nodes in imaging data, imaging and non-imaging data are effectively integrated to achieve comprehensive and accurate AD prediction. In the local learning part, the feature selection process is explained in detail through the Top-k selection strategy and heat map. In the global prediction part, SHAP and LIME are used to perform feature interpretation analysis, which reveals the impact of features on model output, and further verifies the effectiveness of feature selection and the impact of each feature on model decision-making. Compared with classical machine learning and deep learning models, this method performs superiorly, has higher generalization ability and good interpretability.

[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. An Alzheimer's disease diagnosis system based on an interpretable graph neural network, characterized by: include: Preprocessing module: used to preprocess the original neuroimaging data to obtain standard T1 images; Local map construction module: used to segment the standard T1 image into multiple regions of interest to form a local map; Important feature node selection module: It is used to take the local graph as input and select important feature nodes through a self-attention-based pooling mechanism; Node selection process display module: Use the top-k selection strategy in the pooling mechanism combined with the heat map to explain the important feature node selection process and results; Global graph construction module: used to construct data samples by combining selected important node features and non-imaging features, and then randomly select known samples and unknown samples from the data set to form a global graph; Alzheimer's disease diagnosis module: used to use the global graph as the input of the GNN layer, adaptively learn the relationship between nodes, predict Alzheimer's disease, and obtain the diagnosis result of Alzheimer's disease; Global Attribution Explanation Module: Visualize the impact of features on results through interpretable methods to understand diagnostic decision-making behavior.

2. The Alzheimer's disease diagnosis system based on interpretable graph neural network according to claim 1, characterized in that: The preprocessing module preprocesses the original neuroimaging data, specifically including skull removal, cerebellum resection, image segmentation, standardization, and quality monitoring, so as to obtain a standard T1 image.

3. The Alzheimer's disease diagnosis system based on interpretable graph neural network according to claim 1, characterized in that: The local graph construction module first uses the Harvard-Oxford atlas to segment the preprocessed standard T1 image into multiple regions of interest ROIs, each ROI corresponds to a node, to construct a local graph, and simultaneously obtains the feature matrix of each node and the adjacency matrix of the local graph.

4. The Alzheimer's disease diagnosis system based on interpretable graph neural network according to claim 1, characterized in that: The important feature node selection module performs a pooling operation according to the topological structure of the importance of the nodes, that is, selects the nodes using the top-k strategy and discards unimportant nodes.

5. The Alzheimer's disease diagnosis system based on interpretable graph neural network according to claim 1, characterized in that: The node selection process display module combines the top-k strategy with the heat map, and uses the heat map to present the nodes after the scores are assigned and the nodes after the selection, so as to display the node selection process.

6. The Alzheimer's disease diagnosis system based on interpretable graph neural network according to claim 1, characterized in that: The global graph construction module fuses the new imaging node data with the non-imaging data to form a sample set of the global graph, randomly selects α known samples and β unknown samples from the sample set as the input S of the graph neural network module, and uses the input S to establish relationships on the global graph.

7. The Alzheimer's disease diagnosis system based on interpretable graph neural network according to claim 1, characterized in that: The Alzheimer's disease diagnosis module takes the global graph as input and calculates the edge probability matrix and edge weight matrix, specifically including: Calculate the edge probability matrix E using prior knowledge and risk factors: Where k represents the kth risk factor feature, K is the number of risk factor features, and e i,j is the i-th row and j-th column of the edge probability matrix E, representing the edge weights between nodes; Then, the CNN model is used to calculate the similarity of different nodes in the node set, obtain the absolute difference between each feature of the two nodes, and learn the edge weights from the absolute difference to obtain the edge weight matrix in, Represents the edge weight matrix W of the l-th layer network l The weights between nodes in ; This results in a self-metric adjacency matrix with probability constraints: Represents element-wise multiplication; Use the SOFTMAX layer to normalize the matrix and add it to the adjacency operator family after normalization. In order to update the node; D represents the diagonal matrix, which is used to preserve the characteristics of the node itself when updating the node; Use each operator in the operator family to update the node and accumulate the update results G n (V (l) ), and then add the accumulated result to the input node V of this layer (l) Connect and retain the characteristics of the input node while obtaining the output node V of the GNN layer (l+1) =[V (l) , G n (V (l) )]; Finally, after the feature dimension is changed in the last layer, the output of the unknown node is directly input into the SOFTMAX layer to normalize the output; the normalized result is the prediction of the unknown node, thus completing the diagnosis prediction of Alzheimer's disease.

8. The Alzheimer's disease diagnosis system based on interpretable graph neural network according to claim 1, characterized in that: The global attribution explanation module performs global attribution explanation on the trained model through a model-agnostic local proxy interpretable fusion method, that is, SHAP is used to explain the impact of global features on the model output, and the LIME method is used to analyze the impact of each feature on different categories; thereby explaining the impact of each feature on the prediction result and understanding the diagnostic decision-making behavior of the model.

Images