CN118105035A

CN118105035A - A method for detecting mild cognitive impairment based on the combination of brain region and group information

Info

Publication number: CN118105035A
Application number: CN202410242480.6A
Authority: CN
Inventors: 汪洁; 吴洛煌; 匡湖林
Original assignee: Central South University
Current assignee: Central South University
Priority date: 2024-03-04
Filing date: 2024-03-04
Publication date: 2024-05-31

Abstract

The present invention discloses a method for detecting mild cognitive impairment based on the combination of brain region and group information. In recent years, graph neural networks have performed well in the task of detecting mild cognitive impairment. This task can usually extract information from the topological structure of brain regions and group associations. However, existing methods ignore the fusion of information at these two levels to further improve detection accuracy. The present invention proposes a hypergraph neural network that fuses group graphs and brain region graphs. We model the group association information and brain region topological information in the form of a hypergraph to enhance the model's ability to extract high-order information. At the same time, we use the brain region hypergraph of each subject as its node feature in the group hypergraph, retain its complete brain region topological information, and enable the group and brain region information to be aggregated in an end-to-end manner. In addition, since demographic data can also provide information that is useful for detection, we obtain subject association information from it based on the idea of metric learning to enhance the construction of the group hypergraph. Experimental results show that our method can effectively use information at both the group and brain region levels to accurately detect mild cognitive impairment.

Description

Light cognitive impairment detection method based on combination of brain area and group information

Technical Field

The invention belongs to the field of medical image processing, and relates to a mild cognitive impairment detection method based on combination of brain area and group information.

Technical Field

Alzheimer's Disease (AD) accounts for over 60% of all cases of Alzheimer's Disease, and no effective drug or method for treating AD has been found at present, which has become a major cause of disability and death in the elderly. There is also a transitional phase between normal aging and AD, mild cognitive impairment (Mild Cognitive Impairment, MCI), where early intervention in the patient can effectively delay progression. Therefore, the method accurately detects the MCI patient from the normal aging population, provides an important basis for early intervention and control of disease deterioration, and has important clinical and social significance.

Studies have shown that mild cognitive impairment affects the structure and function of the brain of a patient, and at present MCI detection usually employs imaging means such as structural magnetic resonance imaging (Structural Magnetic Resonance Imaging, sMRI), functional magnetic resonance imaging (Functional Magnetic Resonance Imaging, fMRI) and positron emission tomography (Positron Emission Tomography, PET) to observe abnormal changes in the brain of a subject. These imaging techniques are each characterized by: sMRI can acquire morphological information (atrophy, deformation and the like) of the brain of the subject, and the spatial resolution is higher; fMRI mainly measures the blood oxygen level dependent (Blood Oxygenation LEVEL DEPENDENT, BOLD) signals of various parts of the brain to reflect the functional connectivity status of the brain; PET uses the metabolic profile of radiopharmaceuticals in the brain to reflect the active state of various parts of the brain. fMRI has a higher temporal resolution compared to PET and does not require injection of a radiopharmaceutical; the structural atrophy seen in the MCI phase is minor compared to sMRI, while fMRI can provide effective discrimination information from the functional connectivity level. In addition, studies have shown that abnormalities in brain function network occur early in MCI, and fMRI has a good detection effect on early changes in brain function, and is widely used in MCI detection.

FMRI reflects the brain's connectivity pattern and there is rich topology information. Therefore, many researchers use the graph neural network to extract topology information in fMRI to implement MCI detection, and the graph neural network has become a mainstream method of MCI detection. Most MCI detection methods based on graph neural networks can be divided into two types according to a composition mode, and one type of method uses brain areas as nodes of the graph to construct a connection graph of the brain areas; another class of methods uses subjects as nodes in the graph to construct a population association graph of subjects. The first type of method has the main problem that the demographic data (such as gender and age) of the subjects and the association information between the subjects are ignored, and research shows that the inclusion of the demographic data and the association information between the subjects into the MCI detection method can provide information complementary to the imaging, so that the accuracy of detection can be further improved. The second type of method has the main problem of summarizing whole brain information using low-dimensional vectors as a characteristic representation of the subject, and features extracted by this method lose certain brain topology information. In practice, the MCI detection can obtain important detection information from the brain region connection diagram and the population association diagram at the same time, and the two methods can mutually compensate the respective defects by combining.

Therefore, there is a need for a graphic neural network that fuses brain region map and population map information to achieve accurate detection of MCI based on fMRI data and demographic data.

Disclosure of Invention

Aiming at the defects of the prior art, the invention provides a light cognitive impairment detection method based on the combination of a brain region graph and a group graph, which solves the problem of insufficient utilization of the two types of information by the prior method so as to improve the detection accuracy.

The invention provides a mild cognitive impairment detection method based on the combination of a brain region graph and a population graph, which comprises the following steps of:

step 1: preprocessing fMRI data in the data set, mainly comprising magnetization balance, time correction, head motion correction, registration, spatial standardization and template segmentation; wherein magnetization balancing is achieved by removing the first ten volumes of fMRI scans; time correction and head motion correction are standard preprocessing flows for improving data quality; because of the different brain morphologies of different subjects, in order to be comparable, fMRI data of all subjects need to be aligned, mapped into a standard space; since fMRI reflects BOLD signals at the voxel level, in order to obtain BOLD signals at the brain region level, it is necessary to divide the voxels into different brain regions according to a predefined brain map, the BOLD signals of the brain regions being an average of the voxel BOLD signals it contains. After finishing preprocessing fMRI data, each subject has BOLD signals at brain region level and demographic data;

Step 2: and (3) inputting the BOLD signals obtained in the step (1) into a brain region hypergraph construction module designed by us, wherein the module is used for constructing a corresponding brain region hypergraph for each subject according to the BOLD signals of the subjects. The hypergraph of brain area of each subject reflects the functional connection structure of brain, and the main function in the invention is to extract the higher-order topology information of brain area level;

The brain region hypergraph construction module consists of a BOLD signal selection module, a brain region selection module and a brain region hyperedge construction module;

The BOLD signal selection module is used for selecting BOLD signal points with more information based on a minimum Absolute value contraction and selection operator (LASSO) algorithm;

The brain region selection module selects brain regions which have important influence on MCI detection according to the group difference between the normal aging population and the MCI patient population based on Kendel correlation analysis;

The brain region superside construction module comprises: calculating a brain region distance matrix, calculating a brain region incidence matrix and generating a brain region hypergraph;

The input of the brain region distance matrix calculation is brain region and corresponding BOLD signals after the BOLD signal selection and brain region selection module, and the distance between the brain region BOLD signals is calculated;

The input of brain region association matrix calculation is a brain region distance matrix of a subject, and K brain regions nearest to each brain region are configured as1 in the association matrix according to a K nearest neighbor algorithm;

Generating a brain region hypergraph, namely generating the brain region hypergraph for each subject according to the connection relation indicated by the brain region incidence matrix;

Step 3: inputting demographic data of a subject and the brain region hypergraph obtained in the step 2 into a group hypergraph construction module, wherein nodes in the group hypergraph are brain region hypergraphs in the step 2, and constructing based on the imaging distance and the demographic distance of the subject, so as to reflect the association information between the subjects;

The group hypergraph construction module consists of an imaging distance calculation module, a population distance calculation module and a group hyperedge construction module;

The input of the imaging distance calculation module is brain region hypergraph of each subject, and an imaging distance matrix is obtained according to the distance of the brain region hypergraph of each subject.

The demographic distance calculation module is based on a metric learning algorithm, and searches an optimal space for measuring the demographic data distance according to the threshold similarity of the demographic data of the subject, and the main steps of the demographic distance calculation module are initialization, iterative updating and output;

the initialization of the demographic distance calculation module calculates the threshold similarity of demographic data, configures neighbors of a subject based on a K neighbor algorithm, and initializes a projection matrix as an identity matrix;

The iterative updating of the demographic distance calculation module uses a projection matrix to project demographic data into a new metric space, and the neighbor information of the subject is used for optimizing the projection matrix to find an optimal metric space;

The output of the demographic distance calculation module projects demographic data according to an optimized projection matrix, and the demographic distance matrix is configured as the paired distance of the projected demographic data;

The group superside construction module carries out matrix summation on the imaging distance matrix and the demographic distance matrix by a weight coefficient to obtain a distance matrix for constructing a group supergraph, and configures the group superside based on a K neighbor algorithm to form the group supergraph;

Step 4: inputting the brain region hypergraph obtained in the step 2 and the population hypergraph obtained in the step 3 into a hypergraph neural network to extract the characteristics of population and brain region layers;

The hypergraph neural network consists of a group information transmission layer, a brain region convolution layer and a multi-layer perceptron, wherein the group information transmission layer and the brain region convolution layer are based on spectral domain convolution of the hypergraph. The graph convolution operation on the brain region hypergraph can extract the spatial information of the brain region of each subject, and the information transmission on the group hypergraph enables each subject to aggregate the characteristics of other subjects according to the association relation among the subjects. The group information transmission layers and the brain region hypergraph convolution layers are alternately stacked, brain region characteristics are transmitted among subjects with strong relevance after each group information transmission is completed, and then common space information is further extracted by utilizing brain region hypergraph convolution. Transmitting the extracted information into a multi-layer perceptron to obtain logits values of the corresponding classes of each subject;

step 5: inputting logits values of each subject obtained in step 4 into a softmax function to obtain probability values of the subjects belonging to different categories, and converting logits values into probability values according to the following calculation formula:

Wherein p _i represents the probability that the subject is of class i; a logits value representing the subject on category i; k represents the number of categories.

Step 6: and (3) inputting the probability value of the subject on the training set obtained in the step (5) and the corresponding real label into a loss function, and optimizing the parameters of the hypergraph neural network in the step (4) according to the value of the loss function. The loss function is a cross entropy loss function, and is defined as follows:

wherein N is the number of samples of the training set, and y _i is the real label of the samples;

Step 7: and (3) inputting the test set divided in the step (2) into a trained model to obtain a test result of the model.

According to the invention, the brain region hypergraph is used as a node of the population hypergraph, and the population associated information of the subject is incorporated into the hypergraph neural network framework while the topology information of the brain region is reserved, so that accurate MCI detection is realized. Experimental results show that the model can effectively improve the MCI detection performance.

Drawings

FIG. 1 is a schematic flow chart of the method of the present invention.

FIG. 2 is a schematic representation of a method of the present invention.

Fig. 3 is a schematic diagram of a demographic distance calculation module of the method of the present invention.

Fig. 4 is a hypergraph neural network layer schematic of the method of the present invention.

Advantageous effects

The mild cognitive impairment detection model based on the combination of brain region and group information constructed by the invention is an end-to-end deep learning network, and firstly uses a hypergraph to model topology information of the brain region and the group, and the hypergraph has the advantage of better exploring high-order association information in a topology structure compared with a common graph. The demographic distance calculation module designed in the population hypergraph can effectively extract the subject association information based on demographic data by utilizing the threshold similarity to guide the optimization of the metric space. According to the invention, the brain region hypergraph is used as a node of the population hypergraph, and on the basis of retaining the complete brain region topology information, the association information of the population is also incorporated into the model, so that more accurate mild cognitive impairment detection can be realized by utilizing the brain region and the population information.

Detailed Description

As shown in fig. 1, the method for detecting mild cognitive impairment based on the combination of brain area and group information mainly comprises the following steps:

Step 1: preprocessing fMRI data in the data set, and dividing a training set and a testing set;

conventional preprocessing of fMRI data mainly includes magnetization balancing, time correction, head motion correction, registration, spatial normalization and template segmentation; since the magnetic resonance apparatus may not be in a steady state immediately after start-up, the data of the first 10 time points of the scan need to be removed to ensure magnetization balance; fMRI scanning is usually carried out in a layering mode, and the scanning time of different slices has certain difference and needs to be corrected in time; the head movement correction aims at eliminating head movement influence of a subject during scanning so as to improve data quality; because of the different brain morphologies of different subjects, in order to be comparable, fMRI data of all subjects need to be aligned, mapped into a standard space; since fMRI reflects BOLD signals at the voxel level, in order to obtain BOLD signals at the brain region level, it is necessary to divide the voxels into different brain regions according to a predefined brain map, the BOLD signals of the brain regions being an average of the voxel BOLD signals it contains. After finishing preprocessing of fMRI data, each subject had BOLD signals at brain region level and demographic data, and all subjects were randomly divided into training and test sets.

Step 2: constructing a brain region hypergraph according to the data preprocessed in the step 1;

As shown in FIG. 2, the brain region hypergraph construction module consists of three sub-modules, namely a BOLD signal selection module, a brain region selection module and a brain region hyperedge construction module. Because the BOLD signal dimension is generally longer and the number of available subjects is smaller, the use of a complete BOLD signal results in the occurrence of an overfitting phenomenon, and not all brain regions have an effect on the detection of mild cognitive impairment, the BOLD signal selection module and the brain region selection module are used to preserve meaningful BOLD signals and brain regions, thereby improving the quality of construction of brain region hypergraphs. The brain region superside construction module is used for modeling topological information among brain regions of a subject, and high-order association can be better explored by constructing the brain region superside construction module in a hypergraph mode.

The BOLD signal selection module is realized based on an LASSO algorithm. The module iteratively searches for a BOLD time point which has an important influence on all brain areas, and in the ith iteration process, the module selects a BOLD signal of the ith brain area for each subject as a characteristic of the BOLD signal; then fitting the features of all subjects to LASSO, which will output an index of BOLD signals important for the ith brain region; the module then counts the BOLD signals in this index; after the iteration of all brain regions is completed, a final count result of the BOLD signals can be obtained, the larger the count value is, the larger the influence of the BOLD signals in all brain regions is, and the module selects the first k BOLD signals with the larger count value as output.

The brain region selection module is designed based on the principle that the brain region functions of normal aging population and mild cognitive impairment population have inter-group differences, and brain regions with larger inter-group differences have larger influence on mild cognitive impairment detection. Since the features of each subject are two-dimensional matrices composed of brain regions and BOLD signals, the kendel correlation coefficients cannot be directly calculated, the module first pools the BOLD signals of each brain region, calculates the average BOLD signal intensity of the BOLD signals as input, calculates the kendel correlation coefficients corresponding to each brain region, and outputs k brain regions with larger coefficients as output.

The brain region hyperlimit construction module forms the characteristics of the subject by using the selected BOLD signal and the brain region, and constructs a brain region hypergraph based on the characteristics, wherein one brain region hypergraph can be recorded asWhereinIs the hypergraph point set, in the present invention, the selected brain region; epsilon is the edge set; w is the weight of the hyperedge. Hypergraphs typically use a matrix of indicationsDefinition, which is defined as follows:

Wherein the method comprises the steps of Representing a node, e epsilon represents a superside, one superside in the supergraph can be connected with a plurality of nodes, and H reflects the condition that the superside contains nodes. The brain region overedge construction is based on a K nearest neighbor method, H is constructed according to a brain region distance matrix, and the distance calculation mode of brain regions i and j is as follows:

wherein B is the number of time points of the selected BOLD signals, and X is the two-dimensional feature matrix of the selected BOLD signals corresponding to the brain region. In the K-nearest neighbor method, each brain region is regarded as a center node of a superside, and then is connected with K-1 nodes of the nearest neighbor, so that the superside of the brain region is constructed.

Step 3: constructing a group hypergraph according to the demographic data and the brain region hypergraph obtained in the step 2;

The group hypergraph construction module consists of an imaging distance calculation module, a population distance calculation module and a group hyperedge construction module. The association information between subjects is generally composed of two parts, one part is imaging data, and the other part is demographic data corresponding to the imaging distance calculation module and the demographic distance calculation module respectively. The group superside construction module fuses the two part distances and constructs a group superside to generate a group supergraph.

The imaging distance calculation module is used for obtaining imaging distances between each subject and other subjects, and for the subjects i and j, the imaging distance calculation formula is as follows:

Wherein D _img is an imaging distance matrix of the subject, R is the number of brain regions after selection, and X _i is a feature matrix corresponding to the subject i.

The demographic distance calculation module is used to obtain the paired distances between subjects from demographic data, and is based on a large-interval nearest Neighbor algorithm (LARGE MARGIN NEAREST nearest bor, LMNN), as shown in fig. 3. The main idea of the demographic distance calculation module is to use a projection matrix L to calculate demographic dataProjection into the new feature space and use the distance of the new feature space as its demographic distance, its main steps are initialization, iterative updating and output.

In the initialization, the projection matrix L is initialized to an identity matrix while calculating the threshold-type similarity C _demo of the demographic data, the threshold-type similarity calculation formula for the subjects i and j is as follows:

where P is the number of human oral attributes, Is a function that acts on the p-th demographic property to output either 0 or 1. For discrete demographic attributes (e.g., gender, race, etc.),1, When the p-th personal oral property of subjects i and j are the same; for continuous demographic attributes (e.g., age, mental level, etc.),1, When the p-th personal demographic property difference for subjects i and j is less than a given threshold. The module calculates a threshold type similarity C _demo between subjects and defines its most similar K neighbors for each subject according to the similarity size.

In each iterative update, the demographic data is first projected using a projection matrix:

Wherein the method comprises the steps of For projected demographic data, N is the number of subjects. According toThe demographic distance/>, in the new feature space, of the subject can be calculated

Next, according to the neighbor relation defined in the initializing step, assuming that a node i has two types of neighbors with similarity of c ₁ and c ₂(c₁＞c₂), a boundary is formed firstly according to the distance of the farthest neighbor f with similarity of c ₁ and a fixed value mu, and the node with internal similarity smaller than c ₁ is counted into a loss term epsilon _push:

Where K is the number of nodes accounting for the loss term ε _push, μ+D _demo (i, f), the distance of the boundary to node i. The loss term ε _pull is calculated for all neighbors with similarity c ₁ to node i at the same time:

Where C ₁ is the number of neighbors with similarity to node i of C ₁. Similarly, the loss terms ε _push and ε _pull are calculated in the same way for neighbors with a similarity of c ₂. Finally, a coefficient lambda is used to balance the two loss terms to obtain a final loss value

The loss value optimizes the projection matrix L using the same convex optimization method as the original LMNN algorithm.

In the outputting step, the iteratively updated L is used to project the demographic data and return the pair of demographic distance matrices D _demo in the projection space.

The group superside construction module balances and combines D _img output by the imaging distance calculation module and D _demo output by the demographic distance calculation module by a coefficient theta, and the distance matrix D _total for constructing the group supergraph is calculated as follows:

D_total＝D_img+θD_demo，

Similar to constructing brain region hypergrams, the module uses the K-nearest neighbor algorithm to configure population hypergrams from D _total Indicator matrixAnd generating a population hypergraph. At the same time, the module uses the feature X _i of each subject brain region hypergraph as a feature of a node in the population hypergraph.

Step 4: inputting the brain region hypergraph obtained in the step 2 and the group hypergraph obtained in the step 3 into a hypergraph neural network for training;

The hypergraph neural network consists of a group information transmission layer (PMP), a brain region convolution layer (RC) and a multi-layer perceptron (MLP), wherein the group information transmission layer and the brain region convolution layer are both based on spectral domain convolution of the hypergraph. The graph convolution operation on the brain region hypergraph can extract the spatial information of the brain region of each subject, and the information transmission on the group hypergraph enables each subject to aggregate the characteristics of other subjects according to the association relation among the subjects. The group information transmission layers and the brain region hypergraph convolution layers are alternately stacked, brain region characteristics are transmitted between subjects with strong relevance after each group information transmission is completed, then the brain region hypergraph convolution is utilized to further extract common spatial information, and the two network layer structures are shown in figure 4. And transmitting the extracted information into a multi-layer perceptron to obtain logits values of the corresponding classes of each subject.

The group information transfer layer PMP is designed based on hyperspectral domain signal smoothing, and is used for a group hypergraphThe calculation formula of the adjacency matrix A _p on the spectrum domain is as follows:

Wherein the method comprises the steps of For the hyperedge weight matrix of the population hypergraph, all hyperedge weights of the population hypergraph are considered to be the same in the module,Is set as a unitary matrix,Indication matrix for population hypergraph,For the diagonal matrix of the group hypergraph nodes,Is a diagonal matrix of group supersides,AndThe calculation formula of (2) is as follows:

Wherein the method comprises the steps of AndThe spectral domain adjacency matrix A _p of the population supergraph is used to aggregate the features of its neighbor nodes for each subject in the population supergraph, stack the feature matrices of all subjects as a three-dimensional tensor and record asWhere N is the number of subjects, R is the number of brain regions selected, and B is the number of BOLD time points selected. The group information transfer layer PMP may be expressed by the following formula:

Wherein the method comprises the steps of For the three-dimensional tensor formed by stacking all subject feature matrices of the first layer, B ^(l) is the dimension of the brain region feature on the first layer, and the calculation formula of the ReLU function is as follows:

ReLU(x)＝max(0,x)，

Where x is the value of the input ReLU function. PMP first uses the spectral domain adjacency matrix A _p of the population supergraph with the feature matrix of all subjects Performing tensor multiplication, wherein the operation enables each subject to aggregate the characteristics according to the association information of the subjects and other subjects, and similar subjects aggregate the characteristics among each other; the ReLU function is then used to enhance the ability of the PMP layer to characterize the nonlinear relationship.

The brain region convolution layer RC is based on the hypergraph domain convolution theory, and is used for hypergraph of brain regions of each subjectThe convolution formula is expressed as follows:

Wherein the method comprises the steps of For the feature matrix of subject i at layer l,Is a learnable parameter matrix,The effect of (a) is equivalent to that of the adjacency matrix of the hypergraph of brain region, which is marked asFor characterizing topological information between brain regions. The RC uses the same matrix of learnable parameters for all subjects, enabling it to extract similar information for similar brain area topologies while reducing the number of network parameters. Sigma represents a nonlinear activation function, and a ReLU is selected from the modules as the activation function.

The hypergraph neural network alternately stacks PMP and RC layers, aggregates information of similar subjects according to the associated information of the subjects, and then further fuses brain region topological information of each subject into the features in a spectral domain graph convolution mode. By means of the alternate stacking mode, the hypergraph neural network can effectively fuse group information and brain area information to extract characteristics effective for detecting mild cognitive impairment. After the features of each subject are extracted through the PMP and RC layers, the features are input into a multi-layer perceptron to obtain logits values of the corresponding detection categories of the subjects, wherein the calculation formula of the multi-layer perceptron is as follows:

MLP(x)＝Linear₃(ReLU(Linear₂(ReLU(Linear₁(x)))))，

Wherein Linear is a Linear layer, and the calculation formula is: linear (x) =wx+b,

Cross entropy is widely used in various detection and classification tasks, so the invention uses cross entropy loss to calculate loss values and optimize model parameters. Firstly, logits values of each subject output by the multi-layer perceptron are input into a softmax function to obtain probability values of the subjects belonging to different categories, and a calculation formula for converting logits values into probability values is as follows:

Wherein p _i represents the probability that the subject is of class i; A logits value representing the subject on category i; k represents the number of categories. And then inputting the probability value of the subject on the training set and the corresponding real label into a loss function, and optimizing the parameters of the hypergraph neural network according to the value of the loss function. The loss function is a cross entropy loss function, and is defined as follows:

where N is the number of samples of the training set and y _i is the true label of the sample.

In the training process, an Adam optimizer is adopted to optimize parameters of the model, wherein the parameters are beta ₁＝0.9,β₂ =0.999; the initial learning rate is set to 0.01; the maximum epoch during training was set to 300. To avoid overfitting, the weight decay is set to 0.00005.

Step 5: the pattern of the hypergraph neural network for executing detection is similar to the node classification pattern in the graph neural network, and the training set and the test set are constructed in a group hypergraph, so that the trained model is required to be used for calculating again on the data set and extracting the classification result of the sample in the test set, and the detection performance of the model on the test set can be obtained.

Experiment verification

In order to verify the validity of the proposed detection model of the present invention, the following method was used for comparison with the proposed model:

BrainGNN: the method is a graph neural network method based on brain region information, and provides a graph convolution method for brain region perception, the characteristics of each brain region pass through a common linear layer to obtain influence scores, and the graph pooling is carried out to remove brain regions with lower influence scores, so that convolution of each layer focuses on more important brain regions.

CGRL: it partitions the BOLD signal into a plurality of time slices using sliding windows to exploit the spatio-temporal patterns therein. It first acquires a global pattern across the subject using a graph convolution, then performs a graph convolution on each time segment to acquire a local pattern, the global pattern and the local pattern being concatenated together as a feature for detection.

EVGCN it proposes a pairwise associative encoder that encodes demographic data of a subject to obtain its associative information, which contains multiple linear layers to project the demographic data into a potential space and calculate the associative strength to the subject using cosine similarity, but which cannot be used for the construction of an initial graph.

LGGNN: it proposes a self-attention based pooling module in which mutual information loss is additionally used to ensure that the weights of selected and unselected brain regions are significantly different, the method first uses the module and the graph convolution to extract features for each subject, and then reconstructs a population graph based on the extracted features for detection.

HUnet: it is a hybrid architecture of hypergraph neural network and Unet that uses hypergraphs to model the subject's brain region topology information, applies a graph pooling operation to reduce the number of brain regions during encoding, and uses zero padding during decoding to recover brain regions that were deleted during encoding.

DwHGCN: the method uses hypergraph to model brain region topology information, designs a self-adaptive dynamic superside weight updating mechanism, sets the superside weight as a learnable parameter, updates the superside weight when in each back propagation, and enables each brain region to aggregate information of other brain regions which are more similar through dynamic side weight.

BOIT: the method is a transducer architecture, which uses a sliding window to divide a BOLD signal into a plurality of time slices, each time slice is a token, the BOLD signal in the slice is initially embedded, and the attention mechanism is used to acquire information of the cross-time slice.

In the experiments, the Accuracy (ACC), area under ROC curve (AUC), true positive (SEN), true negative (SPE), F1 score (F1-score) were used to evaluate the detection performance of each method. The ACC reflects the ability of the detection model to predict that the correct sample occupies all samples, and its calculation formula is as follows:

The AUC value measures the overall performance of the model under different thresholds and is an important index in detection tasks. SEN reflects the ability of the model to correctly determine that it is a positive example in all actual positive examples, and the higher SEN, the lower the false-negative rate, the following calculation formula:

the SPE reflects the capability of the model for all actual negative examples and can correctly judge the model as the negative example, and the higher the SPE is, the lower the false alarm rate is, and the calculation formula is as follows:

The F1 score is a comprehensive index for the precision rate and the recall rate, and the calculation formula is as follows:

Wherein the method comprises the steps of The Recall is calculated in the same way as SEN.

The experiments were performed using the public data sets ADNI3 and ADNI 2. The proposed method and the comparison method both adopt the same test set division mode, and ten-fold cross validation is used for evaluation on the test set. The experimental results are shown in table 1, with the numbers within () being the standard deviation, the average and standard deviation of all the metrics under ten fold cross validation are reported in table 1.

Table 1: detection performance of the proposed method and comparison method on ADNI3 and ADNI2 datasets

As can be seen from the above table, the mean ACC, AUC, SEN, SPE and F1-score of the proposed method in ADNI3 dataset were 87.98 (6.00), 87.77 (8.45), 87.50 (14.66), 88.67 (8.28) and 86.87 (7.63), respectively, all due to other comparative methods. Our method is superior to BrainGNN and CGRL because both methods ignore the inclusion of subject-related information provided by demographic data into the detection framework; EVGCN and LGGNN are inferior to our method in performance because the former does not consider fusing brain region topology information and population information, and the latter fuses brain region topology information and population association information to some extent, but LGGNN is too simple to process demographic data to construct an accurate population map; our approach is superior to both HUnet and dwHGCN brain region hypergraph neural network-based approaches, again because both types of approaches do not consider the inclusion of population-related information into the model.

The proposed method obtains results of 81.05 (9.94) ACC, 81.57 (9.38) ACC and 83.95 (9.85) ACC, 82.46 (17.26) ACC in eMCI detection task and lMCI detection task of ADNI2, respectively, which are superior to other comparison methods. Note that in the two detection tasks of ADNI2, SEN, SPE and F1 of the proposed method are not optimal, since the available sample size of these two tasks is small, and constructing a population map using hypermaps easily results in connections that are too dense, which to some extent would impair the effect of population map information transfer, but overall the performance of the proposed method is still superior.

In summary, the invention provides a light cognitive impairment detection method based on the combination of brain region and group information, and experimental results show that the method can effectively utilize topology information from brain region and group associated information and perform fusion so as to realize more accurate light cognitive impairment detection.

Claims

1. The mild cognitive impairment detection method based on the combination of brain area and population information is characterized by comprising the following steps:

step 2: constructing a brain region hypergraph and a population hypergraph according to the fMRI data preprocessed in the step 1;

Step 3: inputting the brain region hypergraph and the group hypergraph obtained in the step 2 into a hypergraph neural network for training;

Step 4: and (3) testing the trained hypergraph neural network on the test set divided in the step (1) to obtain a detection result.

2. The method for detecting mild cognitive impairment according to claim 1, wherein in step 2, the brain region hypergraph construction module mainly comprises a BOLD signal selection module, a brain region selection module and a brain region hyperborder construction module. The BOLD signal selection module and the brain region selection module are used for screening and removing redundant BOLD signals and brain regions with smaller influence so as to ensure the construction quality of the subsequent brain region hypergraph. The brain region hyperedge construction module is used for calculating the distance of the BOLD signals between the selected brain regions, and further obtaining the brain region hypergraph of each subject. The BOLD signal selection module is realized based on an LASSO algorithm. The module iteratively searches for a BOLD time point which has an important influence on all brain areas, and in the ith iteration process, the module selects a BOLD signal of the ith brain area for each subject as a characteristic of the BOLD signal; then fitting the features of all subjects to LASSO, which will output an index of BOLD signals important for the ith brain region; the module then counts the BOLD signals in this index; after the iteration of all brain regions is completed, a final count result of the BOLD signals can be obtained, the larger the count value is, the larger the influence of the BOLD signals in all brain regions is, and the module selects the first k BOLD signals with the larger count value as output. The brain region selection module is designed based on the principle that the brain region functions of normal aging population and mild cognitive impairment population have inter-group differences, and brain regions with larger inter-group differences have larger influence on mild cognitive impairment detection. Since the features of each subject are two-dimensional matrices composed of brain regions and BOLD signals, the kendel correlation coefficients cannot be directly calculated, the module first pools the BOLD signals of each brain region, calculates the average BOLD signal intensity of the BOLD signals as input, calculates the kendel correlation coefficients corresponding to each brain region, and outputs k brain regions with larger coefficients as output.

3. The method for detecting mild cognitive impairment according to claim 1, wherein in step 2, the brain region hyperlimit constructing module constructs the characteristics of the subject using the selected BOLD signal and the brain region, and constructs a brain region hypermap based thereon, one brain region hypermap being able to be written asWhereinIs the hypergraph point set, namely the selected brain region; epsilon is the edge set; w is the weight of the hyperedge. Hypergraphs are typically defined by an indicator matrix H, which is defined as follows:

Where v represents a node, e represents a superside, a superside in the supergraph may be connected to multiple nodes, and H reflects the case where the superside contains nodes. The brain region overedge construction is based on a K nearest neighbor method, H is constructed according to a brain region distance matrix, and the distance calculation mode of brain regions i and j is as follows:

4. The method for detecting mild cognitive impairment according to claim 1, wherein in step 2, the group hypergraph construction module comprises an imaging distance calculation module, a demographic distance calculation module and a group hyperedge construction module. The association information between subjects is generally composed of two parts, one part is imaging data, and the other part is demographic data corresponding to the imaging distance calculation module and the demographic distance calculation module respectively. The group superside construction module fuses the two part distances and constructs a group superside to generate a group supergraph.

5. The method for detecting mild cognitive impairment according to claim 1, wherein in step 2, the imaging distance calculation module is configured to obtain an imaging distance between each subject and other subjects, and for subjects i and j, the imaging distance calculation formula is as follows:

Wherein D _img is an imaging distance matrix of the subject, R is the number of brain regions after selection, and X _i is a feature matrix corresponding to the subject i. The demographic distance calculation module is used to obtain the paired distances between subjects from demographic data, and is based on a large-interval nearest Neighbor algorithm (LARGE MARGIN NEAREST nearest bor, LMNN), as shown in fig. 3. The main idea of the demographic distance calculation module is to use a projection matrix L to calculate demographic data Projection into the new feature space and use the distance of the new feature space as its demographic distance, its main steps are initialization, iterative updating and output. In the initialization, the projection matrix L is initialized to an identity matrix while calculating the threshold-type similarity C _demo of the demographic data, the threshold-type similarity calculation formula for the subjects i and j is as follows:

where P is the number of human oral attributes, Is a function that acts on the p-th demographic property to output either 0 or 1. For discrete demographic attributes (e.g., gender, race, etc.),1, When the p-th personal oral property of subjects i and j are the same; for continuous demographic attributes (e.g., age, mental level, etc.),1, When the p-th personal demographic property difference for subjects i and j is less than a given threshold. The module calculates a threshold type similarity C _demo between subjects and defines its most similar K neighbors for each subject according to the similarity size. In each iterative update, the demographic data is first projected using a projection matrix:

The loss value optimizes the projection matrix L using the same convex optimization method as the original LMNN algorithm. In the outputting step, the iteratively updated L is used to project the demographic data and return the pair of demographic distance matrices D _demo in the projection space.

6. The method for detecting mild cognitive impairment according to claim 1, wherein in step 2, the group hyperedge construction module balances and combines D _img output by the imaging distance calculation module and D _demo output by the demographic distance calculation module by a factor θ, and the distance matrix D _total for constructing the group hypergraph is calculated as follows:

D_total＝D_img+θD_demo，

Similar to constructing the hypergraph of brain regions, the module configures an indication matrix of the population hypergraph p from D _total using a K-nearest neighbor algorithm And generating a population hypergraph. At the same time, the module uses the feature X _i of each subject brain region hypergraph as a feature of a node in the population hypergraph.

7. The method for detecting mild cognitive impairment according to claim 1, wherein in step 3, the hypergraph neural network comprises a population information transfer layer (PMP), a brain region convolution layer (RC) and a multi-layer perceptron (MLP), both of which are based on spectral domain convolution of the hypergraph. The graph convolution operation on the brain region hypergraph can extract the spatial information of the brain region of each subject, and the information transmission on the group hypergraph enables each subject to aggregate the characteristics of other subjects according to the association relation among the subjects. The group information transmission layers and the brain region hypergraph convolution layers are alternately stacked, brain region characteristics are transmitted among subjects with strong relevance after each group information transmission is completed, and then common space information is further extracted by utilizing brain region hypergraph convolution. And transmitting the extracted information into a multi-layer perceptron to obtain logits values of the corresponding classes of each subject.

8. The method for detecting mild cognitive impairment according to claim 1, wherein in step 3, the population information transfer layer is designed based on a hyperspectral domain signal smoothing for a population hypergraphThe calculation formula of the adjacency matrix A _p on the spectrum domain is as follows:

ReLU(x)＝max(0,x)，

9. The method for detecting mild cognitive impairment according to claim 1, wherein in step 3, the brain region convolution layer is based on a hypergraph domain convolution theory, for each subject's brain region hypergraphThe convolution formula is expressed as follows:

10. The method for detecting mild cognitive impairment according to claim 1, wherein in step 4, the multi-layered perceptron takes as input information extracted from a plurality of group information transfer layers and brain region convolution layers alternately stacked, and outputs logits values of detection categories corresponding to the subject. The loss function employed is a cross entropy loss function, defined as follows: