Brain disease identification method based on personalized enhanced graph network

Abstract

The invention discloses a brain disease recognition method based on a personalized enhanced graph network, and relates to the technical field of data recognition. Comprising the steps that a mild cognitive impairment MCI data set and a subjective memory impairment SMC data set are constructed based on resting state magnetic resonance imaging signals rs-fMRI, static FCN and dynamic FCN are used for expanding the MCI data set and the SMC data set to obtain a training data set and a test data set, an individualized graph network is constructed, the training data set is input into the individualized graph network for training, and the individualized graph network is obtained. And obtaining a trained personalized graph network, inputting the training data set and the test data set into the trained personalized graph network to generate a tested personalized graph network, and identifying a mild cognitive impairment stage and a subjective dysmnesia stage through the tested personalized graph network. The scale of the data set can be increased, different shapes and disease features can be simulated, and a personalized recognition method is formed.

Description

Technical Field

[0001] The present invention relates to the technical field of data recognition, and in particular to a brain disease recognition method based on a personalized enhanced graph network. Background Art

[0002] With the improvement of living standards and the increase in life expectancy, about 50 million people worldwide were troubled by dementia in 2018, and this figure is expected to reach 152 million by 2050. Alzheimer's disease (AD) is the most common type of dementia (accounting for 60%-80%), and it is a very serious and irreversible brain disorder. As the early stage of AD, mild cognitive impairment (MCI) has a conversion rate of 10%-15% per year, and more than 50% of the conversion rate will turn into AD within 5 years. At the MCI stage, through certain cognitive training and drug treatment, the deterioration process can be delayed or stopped. Once MCI develops into AD, there is currently no effective treatment method, so it is very important to identify and classify MCI before deterioration.

[0003] Due to its mild clinical symptoms, accurately identifying the MCI stage (including early MCI (EMCI) and late MCI (LMCI)) and its early subjective memory complaints (SMCs) is a very challenging task. Current research mainly focuses on using machine learning algorithms based on imaging markers to assist in the classification of MCI and SMCs. In the application of machine learning to classify mild cognitive impairment, the construction of brain networks and feature learning have attracted the attention of most researchers, and these studies have achieved good results. However, the scale of the MCI and SMC datasets is very small, and the number of neurons in the machine learning network for classification tasks is usually thousands, and the overfitting phenomenon often occurs, which limits the classification performance.

[0004] Therefore, providing a brain disease recognition method based on a personalized enhanced graph network to solve the difficulties existing in the prior art is an urgent problem for those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides a brain disease recognition method based on a personalized enhanced graph network, which can increase the scale of the dataset, simulate different shapes and disease characteristics, and form a personalized recognition method.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A brain disease recognition method based on a personalized enhanced graph network, comprising the following steps:

[0008] S1. Construct a Mild Cognitive Impairment (MCI) dataset and a Subjective Memory Complaints (SMC) dataset based on resting-state functional magnetic resonance imaging (rs-fMRI) signals;

[0009] S2. Use static fully convolutional networks (FCNs) and dynamic FCNs to augment the MCI and SMC datasets to obtain training datasets and test datasets;

[0010] S3. Construct a personalized graph network, input the training datasets into the personalized graph network for training, and obtain a trained personalized graph network;

[0011] S4. Input the training datasets and test datasets into the trained personalized graph network to generate a personalized graph network after testing;

[0012] S5. Identify the mild cognitive impairment stage and the subjective memory impairment stage through the personalized graph network after testing.

[0013] Optionally, in S2, the training datasets are a combination of static FCNs and dynamic FCNs constructed with the same labels; the test datasets are the constructed static FCNs.

[0014] Optionally, the static FCN is to construct an FCN using the entire rs-fMRI time series, and the expression of the static FCN is:

[0015]

[0016] where c ij = corr{x(i),x(j)}, corr{·} represents the similarity coefficient between brain regions x(i) and x(j), x(i) and x(j) respectively represent the rs-fMRI time series images of a brain region, and R is the number of segmented brain regions.

[0017] Optionally, the dynamic FCN is to construct an FCN using partial rs-fMRI time series, including:

[0018] Segment the entire rs-fMRI time series into multiple subsequences with a certain window size and step length, calculate a symmetric matrix for each subsequence, and construct a series of dynamic FCN matrices from the subsequence symmetric matrices. The expression is:

[0019]

[0020] where c (1) , c (2) , …, c (D) are the subsequence symmetric matrices, and D is the number of sample subsequences.

[0021] Optionally, the dynamic FCN is regarded as a virtual individual, and the static FCN is regarded as a real individual. The real individuals and virtual individuals form a large-scale training dataset, and the mutual connections between individuals and groups are established to extract features.

[0022] For the above method, optionally, the personalized graph network in S5 is embodied based on the loss function, and the expression is:

[0023]

[0024] where N represents the total number of samples, x i represents the input feature of the i-th sample, y i represents the i-th sample label, represents the difference between the prediction result of sample i passing through the GCN network and the label, and θ i is the personalized loss function.

[0025] It can be seen from the above technical solutions that, compared with the prior art, the present invention provides a brain disease recognition method based on a personalized enhanced graph network, which has the following beneficial effects: 1) The present invention proposes an enhanced FCN composed of a static FCN, a dynamic FCN and their interactions to form a large-scale training dataset, increasing the scale of the dataset; 2) The present invention proposes to establish the interaction connections between each FCN and the population, which can simulate different shapes and disease characteristics to form a personalized recognition method. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention, and those of ordinary skill in the art can also obtain other drawings according to the provided drawings without creative efforts.

[0027] Figure 1 It is a flowchart of a brain disease recognition method based on a personalized enhanced graph network disclosed by the present invention;

[0028] Figure 2 It is a schematic diagram of the principle of the brain disease recognition method disclosed by the present invention;

[0029] Figure 3 It is a schematic diagram of the principle of obtaining the adjacency matrix disclosed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0030] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0031] Referring to Figure 1 and Figure 2 as shown, the present invention discloses a brain disease recognition method based on a personalized enhanced graph network, including the following steps:

[0032] S1. Construct a mild cognitive impairment (MCI) dataset and a subjective memory impairment (SMC) dataset based on resting-state functional magnetic resonance imaging (rs-fMRI) signals;

[0033] S2. Use static fully convolutional networks (FCNs) and dynamic FCNs to augment the MCI and SMC datasets to obtain training datasets and test datasets;

[0034] S3. Construct a personalized graph network, input the training datasets into the personalized graph network for training, and obtain a trained personalized graph network;

[0035] S4. Input the training datasets and test datasets into the trained personalized graph network to generate a personalized graph network after testing;

[0036] S5. Identify the mild cognitive impairment stage and the subjective memory impairment stage through the personalized graph network after testing.

[0037] Further, in S2, the training datasets are a combination of static FCNs and dynamic FCNs constructed with the same labels; the test datasets are the constructed static FCNs.

[0038] Specifically, the total number of samples is N, the number of samples in the training datasets is M, the number of subjects in the validation datasets is N - M, and the dynamic FCN represents an augmented FCN constructed by using a sliding window method. When calculating the static FCN, the window length is maximized to the entire time range.

[0039] Further, let X = {x1, x2, …, x M , …, x N},

[0040] where N represents the total number of samples, and M represents the number of training set samples. A single sample x(1), x(2), …, x(R) respectively represent rs-fMRI time series images of a brain region, R is the number of segmented brain regions, and L represents the rs-fMRI time series length.

[0041] Furthermore, the static FCN constructs the FCN by using the entire rs-fMRI time series in combination with the Pearson correlation coefficient method (PCC). The expression of the static FCN is as follows:

[0042]

[0043] Among them, the expression of the Pearson correlation coefficient method (PCC) is c ij = corr{x(i), x(j)}, where corr{·} represents the similarity coefficient between brain regions x(i) and x(j), x(i) and x(j) respectively represent the rs-fMRI time series images of a brain region, and R is the number of segmented brain regions.

[0044] Furthermore, the dynamic FCN constructs the FCN by using a part of the rs-fMRI time series, including:

[0045] The entire rs-fMRI time series is segmented into multiple subsequences with a certain window size and step length. For X = {x1, x2, …, x M , …, x N}, the training samples (M training samples) among them are segmented to obtain subsequences. The number of subsequences for each sample is D = [(L - l) / s] + 1, where l represents the sliding window size and s represents the moving step length. Therefore, the k-th subsequence is described as: X (k) = After obtaining a series of subsequences, their FCNs are constructed using PCC, and the expression is:

[0046]

[0047] Among them, corr{·} is the correlation coefficient between brain region x (k) (i) and brain region x (k) (j) of the k-th subsequence. A symmetric matrix is obtained by calculating for each subsequence, and a series of dynamic FCN matrices are constructed from the symmetric matrices of the subsequences. The expression is:

[0048]

[0049] Among them, c (1) , c (2) , …, c (D) are the symmetric matrices of the subsequences.

[0050] Furthermore, the dynamic FCN is regarded as a virtual individual, and the static FCN is regarded as a real individual. The real individual and the virtual individual form a large-scale training data set, and the mutual connection between the individual and the group is established to extract features.

[0051] Specifically, the static FCN and the dynamic FCN for the same sample have the same labels. After the data augmentation method, the number of the training data set changes from M to (D + 1)×M. To suppress the noise in the enhanced graph, the mutual connections between individuals and groups are established as follows: for virtual individuals, the internal connections are ignored, and only their connections with real individuals are established; for these virtual individuals corresponding to the same sample, a part of the connections are discarded to maintain their appropriate diversity; for real individuals, only their internal interactions are established. Refer to Figure 3 As shown, these interactions are considered as the adjacency matrix in the GCN, and each row of the feature matrix represents the extracted features of the corresponding FCN. The first N rows of the feature matrix represent the extracted features of the static FCN, and the other rows represent the extracted features of the dynamic FCN. In the adjacency matrix, the coefficients in the first N columns are retained, and the other columns are set to zero.

[0052] The retained coefficients in the adjacency matrix A are calculated as follows:

[0053] A(i,j) = Sim(x j ,x j )×(r G (M h (i),M h (j)) + r E (M h (i),M h )(j))),

[0054] where Sim(x j ,x j ) represents the similarity between the rs-fmri image features (features extracted based on RFE) between individual i and individual i, which means that higher similarity will be given higher weight. M h represents non-image phenotype information. r represents the distance between phenotypic features, r G represents the distance of gender, r E represents the distance of the acquisition device, and r is defined as:

[0055]

[0056] Finally, the similarity measurement is

[0057]

[0058] where ρ is the similarity distance, and σ represents the similarity kernel scale.

[0059] Furthermore, constructing the personalized graph network in S3 includes: using the RFE method to extract the features of each sample, designing the personalized graph network to re - consider individual differences, and using the personalized graph network to complete the classification task.

[0060] Specifically, the propagation rule of the graph network is as follows:

[0061]

[0062] A is the adjacency matrix, H l is the input of the l-th layer, is the degree matrix, σ(·) is the activation function. The proposed personalized graph network contains two layers of networks, and the activation functions are ReLU() and softmax() respectively. The expression is:

[0063]

[0064] Furthermore, the personalized graph network in S5 is embodied based on the loss function, and the expression is:

[0065]

[0066] where, x i represents the input feature of the i-th sample, y i represents the label of the i-th sample, represents the difference between the prediction result of sample i through the GCN network and the label, and θ i is the personalized loss function.

[0067] Furthermore, the expression of θ i is: θ i = 1 + Sim(x i , x j ).

[0068] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A brain disease recognition method based on personalized enhanced graph network, characterized in that: The following steps are involved: S1. Construct mild cognitive impairment (MCI) dataset and subjective memory impairment (SMC) dataset based on resting-state magnetic resonance imaging (rs-fMRI) signals. S2, using static FCN and dynamic FCN to expand the MCI dataset and SMC dataset to obtain training dataset and test dataset; S3, construct a personalized graph network, input the training data set into the personalized graph network for training, and obtain a trained personalized graph network; S4, inputting the training data set and the test data set into the trained personalized graph network to generate a personalized graph network after the test; S5. Identify the mild cognitive impairment stage and subjective memory impairment stage of Alzheimer's disease through the personalized graph network after testing.

2. A method for identifying brain diseases based on personalized enhanced graph network according to claim 1, characterized in that: The training data set in S2 is a combination of static FCN and dynamic FCN constructed using the same labels; the test data set is the constructed static FCN.

3. A method for identifying brain diseases based on personalized enhanced graph network according to claim 2, characterized in that: Static FCN uses the entire rs-fMRI time series to construct FCN. The static FCN expression is: Among them, c ij =corr{x(i),x(j)}, corr{·} represents the similarity coefficient between brain regions x(i) and x(j), x(i) and x(j) represent the rs-fMRI time series images of a brain region respectively, and R is the number of segmented brain regions.

4. The method for identifying brain diseases based on personalized enhanced graph network according to claim 3, characterized in that: Dynamic FCN constructs FCN using part of the rs-fMRI time series, including: The entire rs-fMRI time series is divided into multiple subsequences with certain window sizes and step sizes. A symmetric matrix is ​​calculated for each subsequence, and the subsequence symmetric matrix is ​​constructed into a series of dynamic FCN matrices, which are expressed as: Among them, c (1) ,c (2) ,…,c (D) is the subseries symmetric matrix, and D is the number of sample subsequences.

5. The method for identifying brain diseases based on personalized enhanced graph network according to claim 4, characterized in that: Dynamic FCN is regarded as a virtual individual, and static FCN is regarded as a real individual. Real individuals and virtual individuals form a large-scale training data set, which establishes mutual connections between individuals and groups and then extracts features.

6. The method for identifying brain diseases based on personalized enhanced graph network according to claim 1, characterized in that: The personalized graph network in S5 is based on the loss function, which is expressed as: Where N represents the total number of samples, x i represents the input feature of the i-th sample, y i represents the i-th sample label, Indicates the difference between the prediction result of sample i through the GCN network and the label, θ i is the personalized loss function.

Images