Mci diagnosis system based on multi-scale hypergraph network

Abstract

The application provides a kind of MCI diagnostic system based on multiscale hypergraph network, storage medium and electronic equipment, it is related to neural network technical field.In the present application, first, the fMRI signal of the patient to be diagnosed is preprocessed, and the BOLD time sequence of each ROI region on the functional image is obtained;Then calculate the phase synchronization relationship of ROI region in the spectral domain by the phase locking value technique;And use the graph attention mechanism to dynamically learn and identify the relative importance between different brain regions in the time domain;Then design a multiscale feature fusion mechanism to integrate the time domain and spectral dynamic connection features of fMRI signal, form a time-frequency fusion matrix, and establish a dynamic hypergraph model based on the matrix;Finally, a detection model for mild cognitive impairment is constructed using a hypergraph neural network, and the parameter optimization is carried out through a grid search strategy to improve the MCI diagnosis effect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of neural networks, and particularly relates to a multi-scale hypergraph network-based MCI diagnosis system, a storage medium and an electronic device. BACKGROUND

[0002] Mild Cognitive Impairment (MCI) is a clinical diagnosis syndrome characterized by objective cognitive decline in one or more cognitive domains. MCI is considered as a transitional state between normal aging and dementia, and is crucial for early diagnosis and intervention.

[0003] Functional Magnetic Resonance Imaging (fMRI) as a non-invasive imaging technique, by measuring Blood Oxygen Level Dependent (BOLD) signals, provides a valuable tool for exploring potential imaging biomarkers of MCI-related neurological complications. fMRI technology has been widely used to reveal the clinical phenotype of MCI, providing the possibility for early diagnosis of the disease.

[0004] Although fMRI technology shows potential in the diagnosis of MCI, there are still some challenges in identifying and detecting MCI based on fMRI brain networks through hypergraph modeling:

[0005] (1) Limitations of vector distance: Traditional vector distance measurement methods rely too much on inferred interactions in geometric space, which may not fully capture the dynamic connection patterns and complex interactions between brain regions. This dependency limits the understanding and analysis capabilities of brain network complexity.

[0006] (2) Challenges of multi-scale feature fusion: In the process of multi-scale feature fusion, how to effectively align and optimize features of different scales is a technical problem. This directly affects the learning and representation ability of hypergraph models for brain network features. SUMMARY

[0007] (I) Technical problems solved

[0008] In view of the deficiencies of the prior art, the present application provides a multi-scale hypergraph network-based MCI diagnosis system, a storage medium and an electronic device, which solves the technical problem of early and accurate diagnosis of MCI.

[0009] (II) Technical solutions

[0010] To achieve the above purposes, the present application is realized by the following technical solutions:

[0011] A MCI diagnosis system based on a multi-scale supergraph network, comprising:

[0012] A preprocessing module configured to acquire and preprocess fMRI signals of a patient to be diagnosed to obtain a BOLD time sequence of each ROI region on a functional image;

[0013] A synchronism analysis module configured to evaluate a degree of phase synchronization between paired ROI time signals by using a phase locking value technique to obtain a second-order spectrum matrix;

[0014] An adaptive module configured to calculate a weight coefficient between the paired ROI time signals by using a graph attention mechanism to obtain a time domain matrix after smoothing;

[0015] A fusion module configured to fuse the second-order spectrum matrix and the time domain matrix to obtain a time-frequency fusion matrix;

[0016] An extraction module configured to take the time-frequency fusion matrix as an input of a supergraph neural network to obtain a feature representation of a high-order relationship between different ROI regions;

[0017] A classification module configured to predict a MCI diagnosis result of the patient to be diagnosed based on a final output of the supergraph neural network.

[0018] Preferably, the preprocessing module preprocesses the fMRI signals by using a DPARSF software package, and specifically includes:

[0019] The original functional image of the fMRI signals is removed at the first 10 time points, slice timing correction is performed, the original functional image is excluded if head motion exceeds 2.5 mm or 2.5 degrees, motion correction is performed, statistical parameter mapping is used for co-registration with a T1 weighted image, and segmentation of gray matter, white matter and cerebrospinal fluid is performed, the functional image is mapped to MNI space by using DARTEL, a half maximum Gaussian kernel is used for 4 mm full width smoothing of the functional image to reduce registration variability, and band-pass filtering is performed on the functional image sequence between 0.01 and 0.1 Hz;

[0020] The obtained BOLD time sequence is represented as:

[0021]

[0022] wherein x i represents a BOLD time sequence of an i-th ROI region as an ROI time signal; represents a real set; N represents a number of ROI regions; F de represents a length of the signal.

[0023] Preferably, the synchronism analysis module is specifically configured to:

[0024] The instantaneous phase of the signal is estimated using the Hilbert transform, as follows:

[0025]

[0026] where, Hx(t) represents the Hilbert transform of the signal x(t); PV represents the Cauchy principal value; x(t') represents the value of the original signal at different times during the integration process; t' represents the integration variable;

[0027] The analytic signal phase is defined as:

[0028]

[0029] where, φ(t) represents the instantaneous phase of the signal; arctan is the inverse tangent function;

[0030] The PLV value between the pair of ROI time signals x i and x j is calculated as follows:

[0031]

[0032] where, l e {1, 2,..., T} represents the time period in T, and T is the total number of time periods; represents the imaginary unit; Δφ ij (t l ) = (φ i (t l )- φ j (t l )) represents the instantaneous phase difference of the ROI time signals, which is calculated from the data x i (t i ) of the lth time period of the ROI time signal x l and the data x j (t j ) of the lth time period of the ROI time signal x l , φ i (t l ) and φ j (t l ) are obtained using the Hilbert transform to obtain the phase value;

[0033] The second-order spectral matrix is formed using the synchronization correlation between the ROI time signals x i and x j , and is defined as follows:

[0034]

[0035] Preferably, the adaptive module is specifically used for:

[0036] The weight coefficient e between the ROI time signals x and x is calculated by using the graph attention mechanism i and x j : ij As the connection strength in the time domain:

[0037]

[0038] Where exp is the exponential function; LeakyReLU is the activation function; W e are the parameterized vector and the learnable weight matrix respectively; || is the connection operation; if is greater than the preset threshold τ, represents the node set directly related to x i , and k is the node index in the node set;

[0039] The attention matrix M is defined as represents the correlation of all ROI time signals in the time domain, and the optimal combination of the learnable parameters is obtained by using the minimization optimization function

[0040]

[0041] Where λ≥0 represents the regularization parameter;

[0042] The all-one matrix I is defined as e :

[0043] C TEM = M TS + M TEM (8)

[0044] Where C SPE represents the smoothed time domain matrix.

[0045] Preferably, the fusion module is specifically used for:

[0046] C TS = C TEM ⊙ C SPE (9)

[0047] Where C m represents the time-frequency fusion matrix; C TS , C TS represent the time domain matrix and the bispectrum matrix respectively; and ⊙ represents the Hadamard product.

[0048] Preferably, the extraction module is specifically used for:​​​

[0049] Defining a hypergraph It means that among them Let ε represent a set of nodes, where n is the total number of nodes; ε = {e1, e2, ..., e...} m} represents the set of superedges, where m is the total number of superedges; Let be the weight matrix of the hyperedges; each hyperedge e connects multiple nodes v, and is composed of... Weighted diagonal matrix;

[0050] SuperMap Represented as The correlation matrix H is defined as follows:

[0051]

[0052] Based on the time-frequency fusion matrix C TS Construct a hyperedge, for the i-th ROI region, in C TS Select the associated region K with the highest connectivity value and define it as the set top_K. i ;

[0053] Based on set top_K i The correlation matrix H is redefined as:

[0054]

[0055] Where j represents the set top_K i Index of related regions;

[0056] For an incidence matrix H, the degree matrix of its vertices is defined as follows: use It is represented as follows, where diag is a diagonal matrix, and d(v i ) represents node v i The degree of the hyperedge and its degree matrix are represented as follows: in δ(e i ) represents the superedge e i The degree;

[0057] The hypergraph neural network is used to capture high-order relationships between different ROI regions for feature representation, where the first... Each convolutional layer is represented as:

[0058]

[0059] Where the superscript T indicates transpose; X (l) Indicates the first The convolutional output features of the layer As

[0060] as a filter for extracting node features in the hypergraph, represents a learnable hyperedge weight matrix, W(e i represents a weight coefficient of the hyperedge e i .

[0061] Preferably, the classification module trains a multilayer perceptron to predict the MCI diagnosis result of the patient to be diagnosed, denoted as:

[0062]

[0063] wherein, represents the MCI diagnosis result of the patient to be diagnosed; MLP represents a multilayer perceptron; and Z represents the final output of the hypergraph neural network.

[0064] Preferably, the hypergraph neural network uses a cross-entropy function to estimate the prediction loss in the training stage, denoted as:

[0065]

[0066] wherein, s represents the index of the subject sample; S is the total number of subject samples; y s , respectively represent the true label and the predicted label of the s-th subject sample; and log is a logarithmic function.

[0067] A storage medium storing a computer program for MCI diagnosis based on a multi-scale hypergraph network, wherein the computer program causes a computer to control the MCI diagnosis system as described above to perform MCI diagnosis.

[0068] An electronic device comprising:

[0069] one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, so that the one or more processors control the MCI diagnosis system as described above to perform MCI diagnosis.

[0070] (III) Beneficial effects

[0071] The present application provides an MCI diagnosis system based on a multi-scale hypergraph network, a storage medium and an electronic device. Compared with the prior art, the following beneficial effects are achieved:

[0072] In the present application, firstly, the fMRI signal of the patient to be diagnosed is preprocessed to obtain the BOLD time sequence of each ROI region on the functional image; then the phase synchronization relationship of the ROI region in the frequency domain is calculated through the phase locking value technology; and the relative importance between different brain regions is dynamically learned and identified in the time domain by using the graph attention mechanism; then a multi-scale feature fusion mechanism is designed to integrate the time domain and frequency spectrum dynamic connection features of the fMRI signal to form a time-frequency fusion matrix, and a dynamic hypergraph model is established based on the matrix; finally, a hypergraph neural network is used to construct a detection model for mild cognitive impairment, and a grid search strategy is used for parameter optimization to improve the MCI diagnosis effect. BRIEF DESCRIPTION OF DRAWINGS

[0073] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0074] Figure 1 The structure block diagram of the MCI diagnosis system based on the multi-scale hypergraph network provided by the embodiments of the present application is shown in the figure.

[0075] Figure 2 The flowchart of the MCI diagnosis method based on the multi-scale hypergraph network provided by the embodiments of the present application is shown in the figure. DETAILED DESCRIPTION

[0076] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the technical scheme in the embodiments of the present application is described clearly and completely. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0077] The embodiments of the present application provide a MCI diagnosis system based on a multi-scale hypergraph network, a storage medium and an electronic device, which solve the technical problem of early and accurate diagnosis of MCI.

[0078] The technical scheme in the embodiments of the present application is to solve the above technical problem, and the general idea is as follows:

[0079] For the diagnosis problem of mild cognitive impairment disease, the embodiments of the present application propose a multi-scale hypergraph network system based on synchronization and attention fusion. Compared with the prior art, the accuracy and efficiency of mild cognitive impairment (MCI) detection are significantly improved.

[0080] Specifically, by using the phase-locking value (PLV) and the advanced attention mechanism, the application can accurately capture the phase synchronization relationship between brain regions and the importance of different brain regions in the time domain, and then integrate the time-frequency information through a multi-scale feature fusion mechanism to construct a dynamic hypergraph model. This model not only optimizes the representation of the complex structure and function of the brain network, but also further improves the diagnostic effect through parameter optimization of the hypergraph neural network and the grid search strategy.

[0081] In order to better understand the above technical solutions, the above technical solutions will be described in detail below in combination with the drawings in the specification and specific embodiments.

[0082] Embodiment 1:

[0083] As shown in the Figure 1 , the application embodiment provides an MCI diagnosis system based on a multi-scale hypergraph network, which comprises:

[0084] A preprocessing module is configured to acquire and preprocess the fMRI signal of a patient to be diagnosed to obtain the BOLD time sequence of each ROI region on the functional image;

[0085] A synchronicity analysis module is configured to evaluate the phase synchronization degree between paired ROI time signals by using the phase-locking value technology to obtain a second-order spectrum matrix;

[0086] An adaptive module is configured to calculate the weight coefficient between paired ROI time signals by using a graph attention mechanism to obtain a time domain matrix after smoothing;

[0087] A fusion module is configured to fuse the second-order spectrum matrix and the time domain matrix to obtain a time-frequency fusion matrix;

[0088] An extraction module is configured to take the time-frequency fusion matrix as the input of a hypergraph neural network to obtain the feature representation of the high-order relationship between different ROI regions;

[0089] A classification module is configured to predict the MCI diagnosis result of the patient to be diagnosed based on the final output of the hypergraph neural network.

[0090] The application embodiment promotes the early discovery and intervention of MCI, provides personalized diagnostic support for doctors, accelerates the diagnostic process, and provides a new tool for neuroscience research, which has important clinical and scientific research value.

[0091] As shown in the Figure 2 , Figure 2 a flowchart of an MCI diagnosis method based on a multi-scale hypergraph network is provided. Next, each module of the above scheme will be described in detail. Figure 2

[0092] ​For the preprocessing module, it is used to acquire and preprocess the fMRI signal of the patient to be diagnosed to obtain the BOLD time series of each ROI region on the functional image.

[0093] The module first acquires the original image data of the fMRI signal. The fMRI technology captures brain activity by acquiring magnetic resonance signals on the head of the patient to be diagnosed, and the data is usually stored in DICOM (Digital Imaging and Communications) format.

[0094] Since the original fMRI signal is 4-dimensional, it contains 3-dimensional spatial information and 1-dimensional time. In practical applications, more research and analysis are done using brain image time series, and fMRI signals cannot be directly used for related research. Therefore, the original data also needs to be preprocessed before using the data.

[0095] To obtain the BOLD time series of each ROI region on the functional image, one or more regions of interest (ROI) can be defined in the brain region of interest to complete. Common methods include anatomy-based methods (such as template-based ROI) and function-based methods (such as independent component analysis or seed correlation).

[0096] For example, the module uses the DPARSF software package to preprocess the fMRI signal, which specifically includes:

[0097] Remove the original functional image of the first 10 time points of the fMRI signal, perform slice timing correction; exclude the original functional image with head motion exceeding 2.5 mm or 2.5 degrees, perform motion correction; use Statistical Parametric Mapping (SPM) and T1 weighted image for co-registration, and perform gray matter, white matter and cerebrospinal fluid segmentation; use DARTEL to map the functional image to the MNI (Montreal Neurological Institute) space proposed by the Montreal Neurological Institute; use a Full width at half maximum (FWHM) Gaussian kernel to smooth the functional image to 4 mm full width to reduce registration variability; band-pass filter the functional image sequence between 0.01-0.1 Hz.

[0098] Finally, the obtained BOLD time series is represented as:

[0099]

[0100] Where x i represents the BOLD time series of the i-th ROI region as the ROI time signal. R represents the real number set; N represents the number of ROI regions; F de represents the length of the signal.

[0101] For the synchronism analysis module, it is used to evaluate the degree of phase synchronization between the paired ROI time signals by using the phase locking value technology, and obtain the second order spectrum matrix.

[0102] In detail, as shown in the figure, the module is specifically used for: Figure 2

[0103] The instantaneous phase of the signal is estimated using the Hilbert transform, as follows:

[0104]

[0105] wherein, represents the Hilbert transform of the signal x(t); PV represents the Cauchy principal value; x(t') represents the value of the original signal at different times during the integration process; t' represents the integration variable.

[0106] The analytical signal phase is defined as:

[0107]

[0108] wherein, φ(t) represents the instantaneous phase of the signal; arctan is the inverse tangent function.

[0109] The PLV value between the paired ROI time signals x i and x j is calculated, as follows:

[0110]

[0111] wherein, l∈{1,2,...,T} represents the time period in T, and T is the total number of time periods; represents the imaginary unit; Δφ ij (t l ) = (φ i (t l )-φ j (t l )) represents the instantaneous phase difference of the ROI time signal, which is calculated from the data x i (t i ) of the lth time period of the ROI time signal x l and the data x j (t j ) of the lth time period of the ROI time signal x l , φ i (t l ) and φ j (t l ) are obtained, and the phase value is obtained by using the Hilbert transform.​

[0112] Using ROI time signal x i and x j The synchronous correlation between them forms the second-order spectral matrix. The definition is as follows:

[0113]

[0114] For the adaptive module, it is used to calculate the weight coefficients between pairs of ROI time signals using the graph attention mechanism, and then obtain the time domain matrix after smoothing.

[0115] To elaborate, such as Figure 2 As shown, this module is specifically used for:

[0116] Calculate the ROI time signal x using graph attention mechanism i and x j The weighting coefficient e between ij As a connection strength in the time domain:

[0117]

[0118] Where exp is the exponential function; LeakyReLU is the activation function; W e These are the parameterized vector and the learnable weight matrix, respectively; || is the join operation; if Greater than the preset threshold τ Indicates with x i The set of directly related nodes, where k is the node index in this set.

[0119] use Define the attention matrix This represents the correlation of all ROI time signals in the time domain, and utilizes the minimization optimization function. To obtain the optimal combination of learnable parameters:

[0120]

[0121] Where λ≥0 represents the regularization parameter.

[0122] To prevent excessive focus on specific ROI regions, a matrix of all ones is then defined. Smooth attention matrix M e :

[0123] C TEM =M e +M ones (8)

[0124] Among them, C TEMThis represents the smoothed time-domain matrix.

[0125] The fusion module is used to fuse the second-order spectral matrix and the time-domain matrix to obtain a time-frequency fusion matrix.

[0126] Considering the real changes in actual bioelectrical signals during their contribution to brain networks, this invention focuses on a multi-scale feature fusion mechanism to fuse the temporal and spectral dynamic connectivity features of fMRI.

[0127] To elaborate, C SPE It provides the synchronization relationship of ROI in the spectrum domain, C TEM It provides the importance relationships of ROIs in the time domain. For example... Figure 2 As shown, this module, from a multi-scale perspective, uses a Hadamard product to transform C... TEM and C SPE Combined, this reveals the synchronicity and dynamic changes between ROI regions:

[0128] C TS =C TEM ⊙C SPE (9)

[0129] Among them, C TS Represents the time-frequency fusion matrix; C TEM C SPE represents the time-domain matrix and the second-order spectral matrix, respectively; ⊙ represents the Hadamard product.

[0130] The extraction module is used to take the time-frequency fusion matrix as input to the hypergraph neural network to obtain feature representations of high-order relationships between different ROI regions.

[0131] like Figure 2 As shown, this module utilizes the time-frequency fusion matrix C TS Forming a hypergraph representing features; specifically used for:

[0132] Defining a hypergraph It means that, among them Let ε represent a set of nodes, where n is the total number of nodes; ε = {e1, e2, ..., e...} m} represents the set of superedges, where m is the total number of superedges; Let be the weight matrix of the hyperedges; each hyperedge e connects multiple nodes v, and is composed of... Weighted diagonal matrix.

[0133] SuperMap Represented as The correlation matrix H is defined as follows:

[0134]

[0135] Based on the time-frequency fusion matrix C TS Construct a hyperedge, for the i-th ROI region, in C TS Select the associated region K with the highest connectivity value and define it as the set top_K. i .

[0136] Based on set top_K i The correlation matrix H is redefined as:

[0137]

[0138] Where j represents the set top_K i Index of the associated region.

[0139] For an incidence matrix H, the degree matrix of its vertices is defined as follows: use It is represented as follows, where diag is a diagonal matrix, and d(v i ) represents node v i The degree of the hyperedge and its degree matrix are represented as follows: in δ(e i ) represents the superedge e i The degree.

[0140] The hypergraph neural network is used to capture high-order relationships between different ROI regions for feature representation, wherein the l-th convolutional layer is represented as:

[0141]

[0142] Where the superscript T indicates transpose; X (l) Θ represents the output feature of the convolution in layer l. (l) It is used as a filter to extract node features in the hypergraph. Let W(e) represent the learnable hyperedge weight matrix. i ) represents the superedge e i The weighting coefficients.

[0143] Finally, after extracting features through a hypergraph neural network, the output Z is used for subsequent graph-level reading and disease classification.

[0144] The classification module is used to predict the MCI diagnosis result of the patient to be diagnosed based on the final output of the hypergraph neural network.

[0145] To elaborate, such as Figure 2 As shown, this module trains a multilayer perceptron to predict the MCI diagnosis result of the patient to be diagnosed, which is represented as follows:

[0146]

[0147] wherein, represents the MCI diagnosis result of the patient to be diagnosed; MLP represents a multi-layer perception; and Z represents the final output of the hypergraph neural network.

[0148] It should be particularly pointed out that the hypergraph neural network in the embodiment of the present application selects a cross-entropy function to estimate the prediction loss in the training stage, which is represented as:

[0149]

[0150] wherein, s represents the index of the subject sample; S is the total number of subject samples; y s 、 respectively represent the true label and the predicted label of the s-th subject sample; and log is a logarithmic function.

[0151] So far, the embodiment of the present application introduces the complete execution process of the MCI diagnosis system based on the multi-scale hypergraph network.

[0152] Embodiment 2:

[0153] The embodiment of the present application provides a storage medium which stores a computer program for MCI diagnosis based on a multi-scale hypergraph network, wherein the computer program causes a computer to control the MCI diagnosis system as described in Embodiment 1 to perform MCI diagnosis.

[0154] Embodiment 3:

[0155] The embodiment of the present application provides an electronic device, comprising:

[0156] one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, so that the one or more processors control the MCI diagnosis system as described in Embodiment 1 to perform MCI diagnosis.

[0157] It can be understood that the storage medium and the electronic device provided by the embodiment of the present application correspond to the MCI diagnosis system based on the multi-scale hypergraph network provided by the embodiment of the present application, and the explanation, examples and beneficial effects of the related content can refer to the corresponding part in the MCI diagnosis system. Here, it will not be repeated.

[0158] In summary, compared with the prior art, the following beneficial effects are achieved:

[0159] 1、The embodiment of the present application, first preprocess the fMRI signal of the patient to be diagnosed, obtain the BOLD time sequence of each ROI region on the functional image;Then calculate the phase synchronization relationship of the ROI region in the frequency domain through the phase locking value technology;And use the graph attention mechanism to dynamically learn and identify the relative importance between different brain regions in the time domain;Then design a multi-scale feature fusion mechanism to integrate the time domain and frequency domain dynamic connection features of the fMRI signal, form a time-frequency fusion matrix, and establish a dynamic hypergraph model based on the matrix;Finally, a detection model for mild cognitive impairment is constructed by using the hypergraph neural network, and the parameter optimization is carried out through the grid search strategy to improve the MCI diagnosis effect.

[0160] 2、The embodiment of the present application promotes the early discovery and intervention of MCI, provides personalized diagnostic support for doctors, accelerates the diagnosis process, and provides a new tool for neuroscience research, which has important clinical and scientific value.

[0161] It should be noted that in this paper, relationship terms such as first and second 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 the entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment. Without more limitations, the element defined by the statement "including a" does not exclude the presence of other identical elements in the process, method, article or equipment including the element.

[0162] The above embodiments are only used to illustrate the technical solutions of the present application, but not to limit them;Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part of the technical features;And these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A multi-scale hypergraph network based MCI diagnosis system, characterized in that, The method comprises the following steps: a preprocessing module is used to acquire and preprocess fMRI signals of a patient to be diagnosed to obtain a BOLD time sequence of each ROI region on a functional image; a synchronism analysis module is used to evaluate the degree of phase synchronization between paired ROI time signals by using a phase-locked value technique to obtain a second-order spectrum matrix; an adaptive module is used to calculate weight coefficients between paired ROI time signals by using a graph attention mechanism to obtain a time domain matrix after smoothing; a fusion module is used to fuse the second-order spectrum matrix and the time domain matrix to obtain a time-frequency fusion matrix; an extraction module is used to take the time-frequency fusion matrix as an input of a hypergraph neural network to obtain a feature representation of a high-order relationship between different ROI regions; a classification module is used to predict an MCI diagnosis result of the patient to be diagnosed based on a final output of the hypergraph neural network. The fusion module is specifically configured to: C TS = C TEM O C SPE where C TS denotes the time-frequency fusion matrix; C TEM , C SPE denote the time-domain matrix and the bispectrum matrix, respectively; and denotes the Hadamard product. The extraction module is specifically configured to: Definition of hypergraph is denoted by denotes the set of nodes, n is the total number of nodes; ε = {e1, e2,..., e m} denotes the set of hyperedges, m is the total number of hyperedges; is the weight matrix of the hyperedge; each hyperedge e connects multiple nodes v, and is weighted by the diagonal matrix of hypergraph is represented as a hypergraph H, defined as follows: Based on time-frequency fusion matrix C TS Constructing super-edge, for the i-th ROI region, in C TS Select the associated region K with the highest connection value in C i ; Based on set top_K i The association matrix H is redefined as: wherein j represents the index of the set top_K i the index of the region of interest. For the incidence matrix H, the degree matrix of its vertices is defined as Using where diag is the diagonal matrix, d(v i ) denotes the degree of node v i ; and the degree matrix of its hyperedges is denoted as where δ(e i ) denotes the degree of hyperedge e i ; The supergraph neural network is used to capture high-order relationships between different ROI regions for feature representation, wherein a first convolutional layer is represented as: ​ where the superscript T denotes transpose; X (l) denotes the convolutional output feature of the layer, which is used as a filter to extract node features in the hypergraph, denotes the learnable hyperedge weight matrix, W(e i ) denotes the weight coefficient of the hyperedge e i .

2. The MCI diagnosis system of claim 1, wherein, The preprocessing module is configured to preprocess the fMRI signals by using a DPARSF software package, specifically including: The original functional images of the fMRI signals are removed for 10 time points, slice timing correction is performed, the original functional images with head motion exceeding 2.5 mm or 2.5 degrees are excluded, motion correction is performed, statistical parameter mapping is used for co-registration with a T1 weighted image, and gray matter, white matter and cerebrospinal fluid are segmented, the functional images are mapped to MNI space by using DARTEL, a half maximum Gaussian kernel is used for 4 mm full-width smoothing of the functional images to reduce registration variability, and the functional image sequence is band-pass filtered between 0.01-0.1 Hz; The obtained BOLD time sequence is represented as: where x i represents the BOLD time series of the i-th ROI region as the ROI time signal; represents the set of real numbers; N represents the number of ROI regions; F de represents the length of the signal.

3. The MCI diagnosis system of claim 2, wherein, The synchronism analysis module is specifically configured to: The instantaneous phase of the signal is estimated by using a Hilbert transform, as shown in the following formula: wherein Hx(t) represents the Hilbert transform of the signal x(t); PV represents the Cauchy principal value; x(t') represents the value of the original signal at different times during the integration process; and t' represents the integration variable. The analytical signal phase is defined as: Wherein, φ(t) represents the instantaneous phase of the signal; arctan is an inverse tangent function; The PLV value between pairs of ROI time signals x i and x j is computed as follows: wherein, l∈{1,2,…,T} represents a time period in T, and T is the total number of time periods; represents a virtual unit; Δφ ij l i l j l represents the instantaneous phase difference of the ROI time signal, which is calculated from the data x i (t i ) of the lth time period of the ROI time signal x l (t j ) and the data x j (t l ) of the lth time period of the ROI time signal x i (t l ) and φ j (t l ), and the phase values are obtained by using Hilbert transform;​​​​​ Using the ROI time signal x i and x j between the synchronization correlation, forming the second order spectrum matrix is defined as follows:

4. The MCI diagnosis system of claim 2, wherein, The adaptive module is specifically configured to: The ROI temporal signal x is calculated using a graph attention mechanism i and x j between the weight coefficients e ij as the connection strength in the time domain: Wherein, exp is an exponential function; LeakyReLU is an activation function; W e are a parameterized vector and a learnable weight matrix, respectively; || is a concatenation operation; if is greater than a preset threshold τ, represents a node set directly related to x i k is a node index in the node set. Adopting Defining the matrix of interest Representing the correlation of all ROI time signals in the time domain and using a minimization optimization function Obtaining the best combination of learnable parameters: Wherein, λ≥0 represents a regularization parameter; defining all-ones matrix smoothed attention matrix M e : C TEM = M e + M ones where C TEM denotes the smoothed time-domain matrix.

5. The MCI diagnosis system of claim 1, wherein, The classification module is configured to train a multilayer perceptron to predict the MCI diagnosis result of the patient to be diagnosed, represented as: wherein, represents the MCI diagnosis result of the patient to be diagnosed; MLP represents a multi-layer perceptron; and Z represents the final output of the hypergraph neural network.

6. The MCI diagnostic system according to any one of claims 1 to 5, characterized in that The hypergraph neural network selects a cross-entropy function to estimate a prediction loss in a training stage, represented as: where s represents the index of the subject sample; S is the total number of subject samples; y s , respectively represent the true label and the predicted label of the s-th subject sample; log is the logarithmic function.

7. A storage medium, characterized by The computer program for MCI diagnosis based on a multi-scale hypergraph network is stored, wherein the computer program enables a computer to control the MCI diagnosis system according to any one of claims 1-6 to perform MCI diagnosis.

8. An electronic device, comprising: The computer program for MCI diagnosis based on a multi-scale hypergraph network is stored, wherein the computer program enables a computer to control the MCI diagnosis system according to any one of claims 1-6 to perform MCI diagnosis. The computer program for MCI diagnosis based on a multi-scale hypergraph network is stored, wherein the computer program enables a computer to control the MCI diagnosis system according to any one of claims 1-6 to perform MCI diagnosis. The computer program for MCI diagnosis based on a multi-scale hypergraph network is stored, wherein the computer program enables a computer to control the MCI diagnosis system according to any one of claims 1-6 to perform MCI diagnosis. The computer program for MCI diagnosis based on a multi-scale hypergraph network is stored, wherein the computer program enables a computer to control the MCI diagnosis system according to any one of claims 1-6 to perform MCI diagnosis.

Images