Mild cognitive impairment detection system, electronic device and storage medium

Abstract

The invention provides a mild cognitive impairment detection system, electronic equipment and a storage medium, and relates to the technical field of brain networks. According to the method, the heterogeneity characteristics of the functional structure brain network based on DTI and fMRI can be identified, the effective characteristics of the homogeneous network area are respectively obtained according to different types of nodes and edges, and the biomarkers judged as the heterogeneity area network are obtained at the same time, so that the characteristics of different areas in the brain are comprehensively and effectively integrated; and whole-brain-level feature representation capable of accurately expressing brain network features of the detected person is formed, so that the detection accuracy of a mild cognitive impairment detection system is improved.

Description

Technical Field

[0001] The present invention relates to the field of brain network technology, and in particular to a mild cognitive impairment detection system, an electronic device and a storage medium. Background Art

[0002] Mild cognitive impairment (MCI) is a clinical symptom characterized by cognitive decline and is considered a precursor to Alzheimer's disease. Early detection of MCI is key to timely intervention and prevention of further cognitive decline. In the prior art, brain diseases are generally analyzed and detected through deep learning methods such as graph convolutional neural networks, and the use of neuroimaging to construct brain networks is a prerequisite for such methods.

[0003] However, many graph neural network studies on brain networks lack heterogeneity analysis of brain networks, and the established graph network methods cannot capture the real heterogeneous connection information and simulate the real biological information flow of the brain, resulting in low accuracy of detection results output by existing mild cognitive impairment detection systems. Summary of the invention

[0004] 1. Technical issues to be resolved

[0005] In view of the deficiencies in the prior art, the present invention provides a mild cognitive impairment detection system, an electronic device and a storage medium, which solve the technical problem that the detection results output by the existing mild cognitive impairment detection system have low accuracy.

[0006] (II) Technical solution

[0007] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0008] In a first aspect, the present invention provides a mild cognitive impairment detection system, comprising a brain network processing module and a brain network detection module;

[0009] The brain network processing module is used to obtain the diffusion tensor imaging neuroimage and the functional magnetic resonance imaging neuroimage of the detected person, and to construct a whole brain network according to the diffusion tensor imaging neuroimage and the functional magnetic resonance imaging neuroimage of the detected person;

[0010] The brain network detection module is used to process the whole brain network of the tested person through a pre-trained mild cognitive impairment detection model based on heterogeneous graph attribute segmentation and integration network, and output a mild cognitive impairment detection result;

[0011] The mild cognitive impairment detection model includes a brain network partitioning layer, a graph diffusion layer, an encoder layer, an adaptive feature fusion layer and a classification prediction layer;

[0012] The brain network division layer is used to divide the whole brain network into a left hemisphere brain network, a hemisphere-interbrain network and a right hemisphere brain network; the left hemisphere brain network and the right hemisphere brain network are both homogeneous attribute networks, and the hemisphere-interbrain network is a heterogeneous attribute network;

[0013] The graph diffusion layer is used to perform diffusion processing on the left hemisphere brain network and the right hemisphere brain network respectively, so as to obtain the left hemisphere brain network and the right hemisphere brain network containing potential connection relationships;

[0014] The encoder layer includes a first homogeneous encoder, a second homogeneous encoder and a heterogeneous encoder, wherein the first homogeneous encoder and the second homogeneous encoder are respectively used to extract network features with homogeneous attributes in the left hemisphere brain network and the right hemisphere brain network containing potential connection relationships; the heterogeneous encoder is used to extract network features with heterogeneous attributes in the inter-hemispheric network;

[0015] Adaptive feature fusion layer: used to fuse the network features extracted by the homogeneous encoder and the heterogeneous encoder. For the same nodes that coexist in different networks, they are merged according to their contribution in their respective networks to obtain a brain network representation at the whole brain level.

[0016] The classification prediction layer is used to process the brain network representation at the whole brain level and output the mild cognitive impairment detection results.

[0017] Preferably, the diffusion processing of the image diffusion layer includes:

[0018] The left hemisphere brain network and the right hemisphere brain network were diffused separately using the heat kernel-based graph diffusion technique.

[0019] Preferably, the diffusion processing of the image diffusion layer includes:

[0020] The PageRank graph diffusion technique was used to diffuse the left hemisphere brain network and the right hemisphere brain network separately.

[0021] Preferably, the homogeneous encoder adopts a deep graph convolutional network that can alleviate the over-smoothing phenomenon, and its l-th layer convolution is expressed as:

[0022]

[0023] Among them, σ is the activation function, α l ∈(0,1] is the preset probability start parameter, β l To adjust the learnable weight matrix W l The hyperparameters of is the left or right hemisphere brain network containing potential connections, H (l) is the hidden layer output of the convolution.

[0024] Preferably, the encoding embedding process of the heterogeneous encoder is expressed as:

[0025]

[0026] Among them, R represents the edge type, A r represents the adjacency structure under the characteristic relationship type, W represents the learnable parameter matrix, X is the feature matrix, H I Indicates heterogeneous region encoding.

[0027] Preferably, the adaptive feature fusion layer is specifically used for:

[0028] Merge the learned representation of the encoder layer output network, and the network feature H of the homogeneous attribute of the right hemisphere brain network R and the homogeneity of the left hemisphere brain network. L , directly splice to form a homogeneous region representation of the brain network:

[0029] H homo =CONCAT(H R ,H L )

[0030] For heterogeneous regions encoding H I , all its nodes are also in H homo is encoded once in ; for this node v that exists in both encodings i ∈N m And corresponding to t∈{H homo ,H hetero} means h i t , a node-level adaptive fusion mechanism is designed to merge nodes, and the corresponding adaptive contribution coefficient is calculated as:

[0031]

[0032] Among them, q, W, b are all learnable parameterized vectors, tanh is the activation function, β i t For node h i At t∈{H homo ,H hetero}'s contribution coefficient, is the result after softmax normalization;

[0033] According to the contribution coefficients of these nodes in different codes, they are fused to obtain the brain network representation Z = {z1, z2, ..., z n}, where z i It is expressed as:

[0034]

[0035] Among them, N m Represents the set of nodes in the whole-brain network.

[0036] Preferably, a multi-layer perceptron is used in the classification prediction layer to read out the brain network representation at the whole brain level and output the mild cognitive impairment detection result.

[0037] Preferably, the loss function in the training process of the pre-trained mild cognitive impairment detection model based on heterogeneous graph attribute segmentation and integration network includes:

[0038] For T samples of tested persons, the prediction output of the mild cognitive impairment detection model for the samples is The actual label of the sample is y, and the cross entropy function is used to optimize the model, which is defined as follows:

[0039]

[0040] In a second aspect, the present invention provides a computer-readable storage medium, which is used to store computer instructions, programs, code sets or instruction sets. When the computer-readable storage medium is run on a computer, the computer executes the functions corresponding to the brain network processing module and the brain network detection module in the mild cognitive impairment system as described above.

[0041] In a third aspect, the present invention provides an electronic device, comprising:

[0042] One or more processors; and a computer-readable storage medium, wherein the computer storage medium is used to store computer instructions, programs, code sets or instruction sets, which, when executed on a computer, enable the one or more processors to implement the functions corresponding to the brain network processing module and the brain network detection module in the mild cognitive impairment system as described above.

[0043] (III) Beneficial effects

[0044] The present invention provides a mild cognitive impairment detection system, an electronic device and a storage medium. Compared with the prior art, the present invention has the following beneficial effects:

[0045] The present invention can identify the heterogeneous characteristics of functional structural brain networks based on DTI and fMRI, obtain the effective characteristics of homogeneous network regions according to the different types of nodes and edges, and simultaneously obtain biomarkers of networks determined to be heterogeneous regions, thereby comprehensively and effectively integrating the characteristics of different regions in the brain, forming a whole-brain-level feature representation that can accurately express the brain network characteristics of the tested person, thereby improving the detection accuracy of the mild cognitive impairment detection system. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0047] Figure 1 This is an overall framework diagram of a mild cognitive impairment detection system according to an embodiment of the present invention;

[0048] Figure 2 Schematic diagram of the division of the whole brain network in an embodiment of the present invention. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0050] The embodiments of the present application solve the technical problem of low accuracy of detection results output by existing mild cognitive impairment detection systems by providing a mild cognitive impairment detection system, an electronic device and a storage medium based on heterogeneous graph attribute segmentation and integration networks, thereby achieving accurate detection of mild cognitive impairment.

[0051] The technical solution in the embodiment of the present application is to solve the above technical problems, and the overall idea is as follows:

[0052] Brain network disease detection technology based on neuroimaging plays an important auxiliary role in doctors' diagnosis work, helps to reduce the influence of subjective factors on doctors' diagnosis, and the invention and application of good technology can greatly improve the diagnosis level of hospitals. However, existing studies have ignored the attribute differences between brain regions of interest (ROIs) and the different connection patterns between them. However, heterogeneity is ubiquitous in brain networks. For example, in terms of node attributes, the left cerebral hemisphere is related to language and logical processing, while the right cerebral hemisphere is good at detecting and responding to emotional changes. In addition, hemispheric specialization, as a key feature of the human brain, also known as brain asymmetry, is believed to be related to the early identification of neurological diseases. At present, there is still a lack of graph-based methods that effectively utilize the heterogeneous structure of brain networks. Although some graph-based brain network analysis methods have recognized the existence of heterogeneity in brain networks, these studies often study heterogeneity from the perspective of graph theory, deviating from the essential characteristics of brain network heterogeneity analysis. In addition, as one of the most complex structures of organisms, the brain network structure reflected by diffusion tensor imaging (DTI) technology only reflects the objective structural connectivity between brain regions through the diffusion of water molecules, and its potential connectivity is less revealed. In fact, brain regions present a variety of complex connection patterns, and revealing these patterns is the key to accurately characterizing the state of brain network activity. Therefore, brain network analysis methods based on graph neural networks need to consider the following two issues:

[0053] (1) Establishing heterogeneous brain networks and analysis methods: Current brain network research for MCI detection is mainly based on the homogeneity assumption and lacks heterogeneity analysis of brain networks. As a result, the established graph network methods cannot capture the true heterogeneous connection information and simulate the real biological information flow of the brain.

[0054] (2) Exploring the potential biological connection patterns of brain networks: Current research on the structure of brain networks mainly relies on the objective structural connectivity of the brain revealed by DTI, while ignoring the potential connection patterns in brain networks, resulting in incomplete and subjective modeling of brain networks.

[0055] In order to solve the above problems, the embodiment of the present invention proposes a mild cognitive impairment detection system based on heterogeneous graph attribute segmentation and integrated network, which better captures the heterogeneous information and real information flow in the brain network by mining the potential connections between brain regions in the brain network, and improves the accuracy of mild cognitive impairment disease detection. The present invention first uses DTI and fMRI to construct a structural functional brain network, wherein DTI describes the structure of the brain network, fMRI is used as a node of the network, and the whole brain network is automatically divided into three parts according to the node type of the network and the DTI connection strength, namely, the left hemisphere homogeneity network, the right hemisphere homogeneity network and the inter-hemispheric heterogeneity network. Then, for the homogeneous network that occupies most of the network, the graph diffusion technology is used to mine the potential connection mode, and the three parts of the network correspond to the graph convolutional neural network method with the best adaptability of each part to perform sub-graph embedding representation learning. Finally, the node representations of the brain regions in each region are fused through the node-level adaptive fusion mechanism, and all sub-graph representations are integrated to form a whole-brain level embedding, thereby predicting the detection results of MCI.

[0056] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0057] An embodiment of the present invention provides a mild cognitive impairment detection system, including a brain network processing module and a brain network detection module;

[0058] The brain network processing module is used to obtain the diffusion tensor imaging (DTI) neuroimages and the functional magnetic resonance imaging (fMRI) neuroimages of the tested person (i.e., the diffusion tensor imaging and fMRI neuroimages are both input into the computer), and to construct a whole-brain network based on the diffusion tensor imaging and fMRI neuroimages of the tested person;

[0059] The brain network detection module is used to process the whole brain network of the tested person through a pre-trained mild cognitive impairment detection model based on heterogeneous graph attribute segmentation and integration network, and output the mild cognitive impairment detection result;

[0060] Among them, the mild cognitive impairment detection model based on heterogeneous graph attribute segmentation and integration network includes brain network partitioning layer, graph diffusion layer, encoder layer, adaptive feature fusion layer and classification prediction layer; the processing process of the mild cognitive impairment detection model on the whole brain network is as follows Figure 1 shown.

[0061] The brain network division layer is used to divide the whole brain network into the left hemisphere brain network, the interhemispheric brain network and the right hemisphere brain network; the left hemisphere brain network and the right hemisphere brain network are both homogeneous attribute networks, and the interhemispheric brain network is a heterogeneous attribute network;

[0062] The graph diffusion layer is used to perform diffusion processing on the left hemisphere brain network and the right hemisphere brain network respectively, and obtain the left hemisphere brain network and the right hemisphere brain network containing potential connection relationships. In this layer, the transfer matrix of the left and right brain networks is reconstructed, the potential connection relationships in the network are mined, and the interfering connection relationships are eliminated;

[0063] The encoder layer includes a first homogeneous encoder, a second homogeneous encoder and a heterogeneous encoder, wherein the first homogeneous encoder and the second homogeneous encoder are respectively used to extract network features with homogeneous attributes in the left hemisphere brain network and the right hemisphere brain network containing potential connection relationships, and learn their network representations; the heterogeneous encoder is used to extract network features with heterogeneous attributes in the inter-hemispheric network and learn its network representation.

[0064] Adaptive feature fusion layer: used to fuse the network features learned by the homogeneous encoder and the heterogeneous encoder. For the same nodes that coexist in different networks, they are merged according to their contribution in their respective networks to obtain a brain network representation at the whole brain level.

[0065] The classification prediction layer is used to process the brain network representation at the whole brain level and output the mild cognitive impairment detection results.

[0066] The embodiments of the present invention can identify the heterogeneous characteristics of functional structural brain networks based on DTI and fMRI, obtain effective characteristics of homogeneous network regions according to different types of nodes and edges, and simultaneously obtain biomarkers of networks determined to be heterogeneous regions, thereby comprehensively and effectively integrating the characteristics of different regions in the brain, forming a whole-brain-level feature representation that can accurately express the brain network characteristics of the person being tested, thereby improving the detection accuracy of the mild cognitive impairment detection system.

[0067] The following is a detailed description of the mild cognitive impairment detection model based on heterogeneous graph attribute segmentation and integrated network in combination with technical principles:

[0068] In the brain network construction and division layer, the basic structure of the whole brain network is determined by the structural connection converted from DTI (diffusion tensor imaging) neuroimaging, and the nodes are characterized by the time series signals of fMRI (functional magnetic resonance imaging) as each brain region. By analyzing the connection strength between brain regions in different cerebral hemispheres and dividing the node types according to the functions of brain regions located in different cerebral hemispheres, the brain network of the subjects is described as a heterogeneous brain network with multiple types of nodes and edges. According to the heterogeneity of the brain network, the whole brain network is divided into a homogeneous left hemisphere brain network, a heterogeneous inter-hemispheric network, and a homogeneous right hemisphere brain network.

[0069] In the graph diffusion layer, for subnetworks such as the left and right hemisphere brain networks that have high homogeneity and include large brain areas and connections, two diffusion methods, graph diffusion technology based on heat kernels and personalized PageRank, are used to perform structural denoising and obtain high-order connections, respectively, as potential brain network structural relationships that cannot be reflected by DTI.

[0070] In the encoder layer, dedicated homogeneous convolutional encoders and heterogeneous encoders are designed for all divided brain network subnetworks to encode brain network subnetworks with different attributes.

[0071] In the adaptive feature fusion layer, a node-level adaptive mechanism is used to fuse the embedding representations of corresponding nodes in different subnetworks based on the encoding results of each encoder, and finally all subnetwork encodings are integrated into a whole-brain-level brain network embedding representation.

[0072] In the classification prediction layer, the whole-brain-level brain network representation is used with a graph-level readout function to obtain the final embedded representation, which then enters the classifier to obtain the final MCI recognition result.

[0073] The following is a further description of each layer in the mild cognitive impairment detection model based on heterogeneous graph attribute segmentation and integration network through specific expressions:

[0074] In the brain network partitioning layer, the structural functional brain network of each MCI subject is first input, represented as G = (A, X), where A ∈ R n×d is the brain network adjacency matrix constructed based on DTI neuroimaging, X = {v1,v2,...,v n}∈R n×d is the brain network node feature matrix constructed based on fMRI, each v i Represents the specific brain region node in the brain network. The brain network of each subject is divided into the left hemisphere brain network G L =(A L ,X L ), right hemisphere brain network G R =(A R ,X R ) and the interhemispheric brain network G I =(A I ,X I ), the specific network division is as follows Figure 2 shown.

[0075] In the graph diffusion layer, a homogeneous deep graph convolutional network is used to L =(A L ,X L ) and G R =(A R ,X R) Implement graph diffusion technology to mine potential network connection patterns and remove some noise connections to obtain new adjacency matrix structures. The general graph diffusion process is defined as follows:

[0076]

[0077] in, represents the general transfer matrix, which is initialized as the adjacency matrix of the brain network, i.e., A R or A L Θ k are polynomial coefficients and satisfy the constraints And Θ k ∈[0,1]. In particular, the heat kernel (Heat) and personalized PageRank (PPR) methods are used to realize the diffusion matrix S based on the brain network.

[0078] Based on the PPR graph diffusion process, the symmetric normalized transfer matrix is ​​defined as T sys =D -1 / 2 AD -1 / 2 , and set Where α∈(0,1) represents the transmission probability of random walk. After simplification, the graph diffusion matrix can be expressed as:

[0079] S PPR =α(I n -(1-α)(I n -T sys )) -1

[0080] Among them, I n ∈R n×n is an identity matrix.

[0081] Based on the graph diffusion process of heat kernel, the normalized transfer matrix of random walk is defined as T rw =AD -1 , and set Where t represents the diffusion time. After simplification, the graph diffusion matrix can be expressed as:

[0082]

[0083] Furthermore, based on the above two graph diffusion matrices, they are regularized respectively, and the final diffused brain network adjacency matrix representation is obtained, that is:

[0084]

[0085] The above two forms of graph diffusion processes based on PPR and heat kernel have the ability to filter out noise connections in the network and mine potential network connections. It should be noted that this diffusion method is only for GL =(A L ,X L ) and G R =(A R ,X R ) was implemented.

[0086] In the encoder layer, for the left hemisphere brain network G L =(A L ,X L ), right hemisphere brain network G R =(A R ,X R ) and the interhemispheric brain network G I =(A I ,X I ) The three networks have two encoders, namely homogeneous encoder and heterogeneous encoder. The homogeneous encoder is mainly used to encode G L =(A L ,X L ) and G R =(A R ,X R ), the heterogeneous encoder is used to encode G I =(A I ,X I ). The homogeneous encoder adopts a deep graph convolutional network that can alleviate the over-smoothing phenomenon. Its l-th layer convolution is expressed as:

[0087]

[0088] Among them, σ is the activation function, α l ∈(0,1] is the preset probability start parameter, β l To adjust the learnable weight matrix W l The hyperparameters of There are two optional forms: PPR and Heat. (l) is the hidden layer output of the convolution. Therefore, for G L =(A L ,X L ) and G R =(A R ,X R ) can be encoded as:

[0089]

[0090] For G I =(A I ,X I ) This type of heterogeneous network designs a heterogeneous encoder for encoding embedding:

[0091]

[0092] Among them, R represents the type of edge, A r represents the adjacency structure under the characteristic relationship type, W represents the learnable parameter matrix, and X is the feature matrix.

[0093] In the adaptive feature fusion layer, the learned representation of the network obtained by the above encoding is merged. R and G L , directly splice to form a homogeneous region representation of the brain network:

[0094] H homo =CONCAT(H R ,H L )

[0095] For heterogeneous regions encoding H I , all its nodes are also in H homo For this node v that exists in both encodings, i ∈N m And corresponding to t∈{H homo ,H hetero} A node-level adaptive fusion mechanism is designed to merge nodes, and the corresponding adaptive contribution coefficient is calculated as:

[0096]

[0097] Among them, q, W, b are all learnable parameterized vectors, tanh is the activation function, β i t is node h i At t∈{H homo ,H hetero}'s contribution coefficient, This is the result after softmax normalization.

[0098] According to the contribution coefficients of these nodes in different codes, they are fused and finally the whole brain network representation Z = {z1,z2,...,z n}, where z i It is expressed as:

[0099]

[0100] In the classification prediction layer, a multi-layer perceptron (MLP) is used to read out the predicted probability for each Z.

[0101] It should be noted that the pre-trained mild cognitive impairment detection model based on heterogeneous graph attribute segmentation and integrated network needs to be trained and optimized before use to determine the parameters of the model. The loss functions in the training process include:

[0102] For T samples of tested persons, the prediction output of the mild cognitive impairment detection model for the samples is The actual label of the sample is y, and the cross entropy function is used to optimize the model, which is defined as follows:

[0103]

[0104] The following verification experiment verifies the effectiveness of the mild cognitive impairment detection model based on heterogeneous graph attribute segmentation and integrated network proposed in the embodiment of the present invention:

[0105] The ADNI dataset was used in the validation experiment. During this validation process, a total of 179 test subject samples were collected, including 28 healthy control (NC) subjects, 51 mild cognitive impairment (MCI) subjects, 64 early mild cognitive impairment (EMCI) subjects, and 36 late mild cognitive impairment (LMCI) subjects.

[0106] During the verification process, the mild cognitive impairment detection model based on heterogeneous graph attribute segmentation and integrated network of the embodiment of the present invention was compared with existing benchmark methods, including support vector machine SVM, GCN, GAT, GCNII, GDC, HebrainGNN, GCNH, and CAGNN.

[0107] Among them, Support Vector Machine: This supervised learning algorithm solves the classification problem by finding the optimal hyperplane with the largest margin in the feature space.

[0108] GCN: Simplifies polynomial convolution into a first-order form to achieve information aggregation within a one-hop neighborhood.

[0109] GAT: Adopts self-attention mechanism to determine the importance of nodes within the first-order neighborhood and promotes local convolution.

[0110] GCNII: is an improved version of the GCN model that combines the initial residual and identity mapping techniques to achieve the stacking of multiple convolutional layers.

[0111] GDC: is a spatially local convolution inspired by spectral methods. It addresses the problems of noisy edges and arbitrary definitions in real graphs by sparsifying the generalized form of graph diffusion.

[0112] HebrainGNN: In this method, the brain network is modeled as a heterogeneous graph, and a white-supervised pre-training strategy is designed to address the overfitting problem caused by the conflict between large-scale parameters and small-scale medical data samples.

[0113] GCNH: is a graph learning method that can be applied to both homogeneous and heterogeneous graphs. It learns and combines the individual representations of nodes and their neighborhoods, and uses learnable coefficients to balance the contributions of central nodes and neighborhoods.

[0114] CAGNN: is an algorithm designed specifically for heterogeneous networks. It decouples node features into discriminative features for downstream tasks, aggregates features for graph convolutional layers, and fuses these two features for classification using a shared mixer module.

[0115] This verification experiment uses four evaluation indicators (accuracy, F1 score, sensitivity and precision) to prove that the mild cognitive impairment detection model based on heterogeneous graph attribute segmentation and integrated network is superior to the above baseline methods in mild cognitive impairment (MCI) disease classification. In order to make a fair comparison, all methods are evaluated by five-fold cross validation, and Bayesian parameter search is performed to find the best parameters. It should be noted that because the graph diffusion mode includes two diffusion modes, heat kernel diffusion and personalized Pagerank, the mild cognitive impairment detection model based on heterogeneous graph attribute segmentation and integrated network has two modes, PageRank-based mode and Heat-based mode, respectively denoted as: HGASIN-PPR, HGASIN-Heat. The specific results are shown in Tables 1 to 4.

[0116] Table 1 Classification results obtained by 5-fold cross validation with competing methods when comparing NC and MCI

[0117]

[0118]

[0119] Table 2 Classification results obtained by 5-fold cross validation when comparing NC number LMCI with competing methods

[0120]

[0121] Table 3 Classification results of NC and EMICI by 5-fold cross validation with competing methods

[0122]

[0123]

[0124] Table 4 Classification results obtained by 5-fold cross validation with competing methods when MCI, LMCI and EMCI are considered as one disease category

[0125]

[0126] In Tables 1 to 3, the ADMI dataset is divided into three tasks: NC vs. MCI, NC vs. LMCI, and NC vs. EMICI, to confirm that the mild cognitive impairment detection model based on heterogeneous graph attribute segmentation and integrated networks can distinguish various fine-grained MCI diseases. Experiments show that in these three discriminative tasks, our two methods have achieved the best results on almost all evaluation indicators. In Table 4, MCI, LMCI, and EMCI are regarded as one disease category and classified together with NC, aiming to verify the competitiveness of the HGASI method in coarse-grained MCI disease detection. Experimental results show that the process based on HGASIN-Heat is more competitive in this task and shows an overall leading effect. Compared with the second-best model GCNH, it improves by 6.73% 6.12.17%, 6.73% and 20.48% in accuracy, F1 value, sensitivity and precision, respectively. In summary, the experimental results in the four tables fully demonstrate the advantages of the mild cognitive impairment detection model based on heterogeneous graph attribute segmentation and integrated network of the embodiment of the present invention in the mild cognitive impairment (MCI) disease detection task, and the heat-based mild cognitive impairment detection model based on heterogeneous graph attribute segmentation and integrated network is superior to the PPR-based mild cognitive impairment detection model based on heterogeneous graph attribute segmentation and integrated network in most cases.

[0127] An embodiment of the present invention also provides a computer-readable storage medium, which is used to store computer instructions, programs, code sets or instruction sets. When the computer-readable storage medium is run on a computer, it enables the computer to perform functions corresponding to the brain network processing module and the brain network detection module in the above-mentioned mild cognitive impairment system.

[0128] An embodiment of the present invention further provides an electronic device, including:

[0129] One or more processors; and a computer-readable storage medium, wherein the computer storage medium is used to store computer instructions, programs, code sets or instruction sets, which, when executed on a computer, enable the one or more processors to implement the functions corresponding to the brain network processing module and the brain network detection module in the above-mentioned mild cognitive impairment system.

[0130] In summary, compared with the prior art, the present invention has the following beneficial effects:

[0131] 1. The embodiments of the present invention can identify the heterogeneous characteristics of the functional structural brain network based on DTI and fMRI, obtain the effective characteristics of the homogeneous network areas according to the different types of nodes and edges, and obtain the biomarkers of the heterogeneous regional network at the same time, so as to comprehensively and effectively integrate the characteristics of different regions in the brain, and form a whole-brain level feature representation that can accurately express the brain network characteristics of the tested person, thereby improving the detection accuracy of the mild cognitive impairment detection system.

[0132] 2. The embodiments of the present invention respectively utilize the methods based on heat kernel and personalized PageRank to denoise the connections between brain region nodes and discover hidden biological connections based on the objective physiological connection structure based on DTI, thereby making up for the lack of effective patterns in the network, so that the mild cognitive impairment system can be more comprehensively and accurately encoded into the characteristics of the brain network.

[0133] 3. The embodiment of the present invention can adaptively fuse the coded representations of nodes from different brain regions. The same brain region node may exist in subnets of different properties. Adaptive mechanisms are used to fuse these node representations and integrate them into the embedding of the whole brain network according to the contributions of these nodes in different subnets. Compared with previous methods, the richer properties of some important nodes are fully considered, thereby further improving the detection accuracy of the mild cognitive impairment detection system.

[0134] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0135] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A mild cognitive impairment detection system, characterized in that: It includes a brain network processing module and a brain network detection module; The brain network processing module is used to obtain the diffusion tensor imaging neuroimage and the functional magnetic resonance imaging neuroimage of the detected person, and to construct a whole brain network according to the diffusion tensor imaging neuroimage and the functional magnetic resonance imaging neuroimage of the detected person; The brain network detection module is used to process the whole brain network of the tested person through a pre-trained mild cognitive impairment detection model based on heterogeneous graph attribute segmentation and integration network, and output a mild cognitive impairment detection result; The mild cognitive impairment detection model includes a brain network partitioning layer, a graph diffusion layer, an encoder layer, an adaptive feature fusion layer and a classification prediction layer; The brain network division layer is used to divide the whole brain network into a left hemisphere brain network, a hemisphere-interbrain network and a right hemisphere brain network; the left hemisphere brain network and the right hemisphere brain network are both homogeneous attribute networks, and the hemisphere-interbrain network is a heterogeneous attribute network; The graph diffusion layer is used to perform diffusion processing on the left hemisphere brain network and the right hemisphere brain network respectively, so as to obtain the left hemisphere brain network and the right hemisphere brain network containing potential connection relationships; The encoder layer includes a first homogeneous encoder, a second homogeneous encoder and a heterogeneous encoder, wherein the first homogeneous encoder and the second homogeneous encoder are respectively used to extract network features with homogeneous attributes in the left hemisphere brain network and the right hemisphere brain network containing potential connection relationships; the heterogeneous encoder is used to extract network features with heterogeneous attributes in the inter-hemispheric network; Adaptive feature fusion layer: used to fuse the network features extracted by the homogeneous encoder and the heterogeneous encoder. For the same nodes that coexist in different networks, they are merged according to their contribution in their respective networks to obtain a brain network representation at the whole brain level. The classification prediction layer is used to process the brain network representation at the whole brain level and output the mild cognitive impairment detection results.

2. The mild cognitive impairment detection system according to claim 1, characterized in that: The diffusion processing of the image diffusion layer includes: The left hemisphere brain network and the right hemisphere brain network were diffused separately using the heat kernel-based graph diffusion technique.

3. The mild cognitive impairment detection system according to claim 1, characterized in that: The diffusion processing of the image diffusion layer includes: The PageRank graph diffusion technique was used to diffuse the left hemisphere brain network and the right hemisphere brain network separately.

4. The mild cognitive impairment detection system according to any one of claims 1 to 3, characterized in that: The homogeneous encoder adopts a deep graph convolutional network that can alleviate the over-smoothing phenomenon, and its l-th layer convolution is expressed as: Among them, σ is the activation function, α l ∈(0,1] is the preset probability start parameter, β l To adjust the learnable weight matrix W l The hyperparameters of is the left or right hemisphere brain network containing potential connections, H (l) is the hidden layer output of the convolution.

5. The mild cognitive impairment detection system according to any one of claims 1 to 3, characterized in that: The expression of the encoding embedding process of the heterogeneous encoder is: Among them, R represents the type of edge, A r represents the adjacency structure under the characteristic relationship type, W represents the learnable parameter matrix, X is the feature matrix, H I Indicates heterogeneous region encoding.

6. The mild cognitive impairment detection system according to any one of claims 1 to 3, characterized in that: The adaptive feature fusion layer is specifically used for: Merge the learned representation of the encoder layer output network, and the network feature H of the homogeneous attribute of the right hemisphere brain network R and the homogeneity of the left hemisphere brain network. L , directly splice to form a homogeneous region representation of the brain network: H homo =CONCAT(H R ,H L ) For heterogeneous regions encoding H I , all its nodes are also in H homo is encoded once in ; for this node v that exists in both encodings i ∈N m And corresponding to t∈{H homo ,H hetero } A node-level adaptive fusion mechanism is designed to merge nodes, and the corresponding adaptive contribution coefficient is calculated as: Among them, q, W, b are all learnable parameterized vectors, tanh is the activation function, For node h i At t∈{H homo ,H hetero }'s contribution coefficient, is the result after softmax normalization; According to the contribution coefficients of these nodes in different codes, they are fused to obtain the brain network representation Z = {z1, z2, ..., z n }, where z i It is expressed as: Among them, N m Represents the set of nodes in the whole-brain network.

7. The mild cognitive impairment detection system according to any one of claims 1 to 3, characterized in that: The classification prediction layer uses a multi-layer perceptron to read out the brain network representation at the whole brain level and output the mild cognitive impairment detection result.

8. The mild cognitive impairment detection system according to any one of claims 1 to 3, characterized in that: The loss function in the training process of the pre-trained mild cognitive impairment detection model based on heterogeneous graph attribute segmentation and integration network includes: For T samples of tested persons, the prediction output of the mild cognitive impairment detection model for the samples is The actual label of the sample is y, and the cross entropy function is used to optimize the model, which is defined as follows:

9. A computer-readable storage medium, characterized in that: The computer storage medium is used to store computer instructions, programs, code sets or instruction sets. When it runs on a computer, it enables the computer to perform the functions corresponding to the brain network processing module and the brain network detection module in the mild cognitive impairment system as described in any one of claims 1 to 8.

10. An electronic device, characterized in that: include: one or more processors; And a computer-readable storage medium, which is used to store computer instructions, programs, code sets or instruction sets, which, when running on a computer, enable the one or more processors to implement the functions corresponding to the brain network processing module and the brain network detection module in the mild cognitive impairment system as described in any one of claims 1 to 8.

Images