Depression brain function connection data analysis method based on multi-map fusion

Abstract

The invention discloses a depression brain function connection data analysis method based on multi-map fusion, is applied to the field of deep learning and nerve imaging, and aims to solve the problem that depression brain function connection cannot be represented accurately and effectively in the prior art. According to the method, brain region level blood oxygen concentration dependency signals are obtained from resting state functional magnetic resonance imaging data based on a plurality of brain maps divided according to different scales or functions, and then a functional connection matrix of each map is obtained through calculation based on a correlation algorithm; and then abstracting the functional connection matrix into graph structure data, carrying out feature extraction by using a multi-scale graph convolutional layer, and finally fusing and analyzing multi-map features by using two different strategies of multi-head cross attention fusion and adaptive map weight learning, so as to execute downstream tasks such as diagnosis or biomarker positioning. According to the method, multi-map information of functional magnetic resonance imaging is well integrated, the reliability of a depression brain function connection analysis result is improved, and an objective reference is favorably provided for clinic.

Description

Technical Field

[0001] The present invention belongs to the fields of deep learning and neuroimaging, and particularly relates to a brain functional connectivity analysis technology for depression. Background Art

[0002] Depression is a common mental illness characterized by a series of cognitive and physical symptoms, including persistent sadness, loss of interest or pleasure, and anhedonia. This disease affects approximately 350 million people globally, imposing a significant burden on society. Approximately half of the patients will have a high risk of recurrence after experiencing their first episode of depression. Therefore, early analysis, diagnosis, and treatment of depression are urgent tasks in the field of mental health. Traditionally, the analysis and diagnosis of depression rely on clinicians to evaluate symptoms through conversations and behavioral observations, while referring to the Diagnostic and Statistical Manual of Mental Disorders, 5th Edition (DSM-5). However, this subjective assessment method sometimes leads to inconsistent clinical diagnosis and treatment results. To overcome this problem, in recent years, neuroimaging techniques such as fMRI have begun to be used to explore the brain functional and structural abnormalities of various mental illnesses. For example, Resting-state functional magnetic resonance imaging (rs-fMRI) helps researchers and clinicians more objectively evaluate the brain's functional connectivity and activity patterns by analyzing the brain's activity in a non-task state, providing a new analysis tool for a wide range of mental illnesses. This technique can not only reveal the neurobiological changes associated with mental illnesses such as depression but also contribute to the development of more precise and personalized treatment plans.

[0003] Due to the high heterogeneity of depression among different individuals, its symptom range is extensive. Different patients may exhibit different symptom combinations, and the severity of each symptom also varies. In addition, the brain structure and function of depression patients show significant individual differences, and compared with other neurological diseases (such as Alzheimer's disease or Parkinson's disease), the brain functional or structural abnormalities are often not obvious enough in images, making the analysis work such as diagnosis for depression extremely challenging.

[0004] With the development of computer technology and deep learning, graph neural networks (GNNs) have been widely used in functional connectivity matrix (FCM) analysis because of their inherent graph structure characteristics that match those of FCMs. GNNs can analyze abnormal connection patterns within FCMs, thus providing support for disease diagnosis and prognosis analysis. However, most existing studies are designed based on a single brain atlas. Since the signal of each brain region is usually calculated as the average blood oxygenation level-dependent (BOLD) signal of all voxels within it, simple averaging may mask these specific signals, resulting in information loss. Secondly, anatomically defined regions of interest (ROIs) may not fully match an individual's brain structure due to inter-individual heterogeneity, which may lead to boundary effects, i.e., the signals of ROI boundary voxels may confound the influence of adjacent brain regions. Summary of the Invention

[0005] In order to accurately and effectively analyze the brain functional connectivity of depression, the present invention proposes a data analysis method for brain functional connectivity of depression based on multi-atlas fusion.

[0006] The technical solution adopted by the present invention is as follows: A data analysis method for brain functional connectivity of depression based on multi-atlas fusion, comprising:

[0007] S1. Collect rs-fMRI data of patients and healthy control groups; as the training data set;

[0008] S2. Preprocess the rs-fMRI data;

[0009] S3. For the preprocessed rs-fMRI data, obtain three groups of BOLD signals at different scales according to the automated anatomical labeling atlas, Harvard-Oxford atlas, and Craddock200 atlas;

[0010] S4. For the three groups of BOLD signals at different scales obtained in step S3, use the Pearson correlation algorithm to calculate and obtain their respective corresponding functional connectivity matrices;

[0011] S5. Represent the functional connectivity matrix as graph structure data;

[0012] S6. Use the graph structure data corresponding to each rs-fMRI data obtained in step S4 from the training data set as the input of the multi-atlas graph representation learning network, and use the patients or healthy control groups corresponding to each rs-fMRI data in the training data set as data labels to train the multi-atlas graph representation learning network;

[0013] S7. Input the three functional connectivity matrices corresponding to the rs-fMRI data to be analyzed into the trained multi-atlas graph representation learning network to obtain the analysis result.

[0014] Advantages of the present invention: The method of the present invention combines the brain functional connectivity information of multiple atlases of rs-fMRI of the subjects, uses the graph neural network method to capture and integrate the structural information and feature information between brain regions, and further enhances the information interaction between atlases by using multi-head cross-attention fusion (MHCAF) and adaptive atlas-weighted learning (AAWL) to comprehensively represent the brain functional connectivity, and finally achieves the purpose of analyzing the brain functional connectivity of depression. It can be used as an objective clinical diagnosis reference to reduce subjective bias and assist doctors in making more efficient and accurate judgments. Description of the Drawings

[0015] Figure 1 It is the overall network flow chart of the present invention.

[0016] Figure 2 It is the method architecture diagram of the present invention. Detailed Embodiments

[0017] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention.

[0018] As Figure 1 , 2 shown, the present invention discloses a method for analyzing brain functional connectivity data of depression based on multi-atlas fusion. Figure 1 It is the flow chart of the present invention. Figure 2 It is the specific method framework diagram. The implementation process of the method of the present invention includes the following steps:

[0019] 1. Collect the rs-fMRI data of patients and healthy controls, and perform data preprocessing operations, including: (1) Discard the first 10 volumes of BOLD signals to reduce initial signal instability; (2) Head motion correction to correct the influence of the subject's head movement during scanning; (3) Spatial normalization to make the images conform to the standard anatomical space; (4) Smoothing to improve the signal-to-noise ratio; (5) Band-pass filtering (0.01 - 0.1 Hz) to isolate the frequency range related to resting-state fluctuations. Immediately extract the BOLD signals of the corresponding ROIs based on multiple brain atlases predefined according to different spatial scales or structural criteria; in this embodiment, three commonly used brain atlases are used as examples for illustration, specifically including: Automated Anatomical Labeling (AAL) atlas, Harvard-Oxford (Harvard) atlas, and Craddock 200 (CC200) atlas. Theoretically, when the computing resources are sufficient, the method of the present invention can process and fuse any number of atlas information. In this embodiment, the above three commonly used atlases are selected, which can comprehensively cover different spatial scales from anatomy to function while ensuring computational efficiency and research interpretability, and are consistent with the widely used research standards in the field;

[0020] Obtain three groups of BOLD signals at different scales from the original input rs-fMRI images according to three predefined atlases represents the real number field, where T is the number of time points of rs-fMRI acquisition of the current subject. For the BOLD signals of each atlas, use the Pearson correlation algorithm (PCC, Pearson correlation coefficient) to calculate and obtain the corresponding functional connectivity matrix M, and the specific calculation is as follows:

[0021] M n (i,j) = corr(b i ,b j ),(i,j = 1,2,…,R n )

[0022] Where corr represents the Pearson correlation algorithm, b i and b j represent the original time series signals of the i-th and j-th brain regions predefined on the n-th atlas, and a total of R n brain regions are predefined on the current atlas.

[0023] 2. Represent the constructed functional connectivity matrix as graph structure data, that is, generate a feature matrix and an adjacency matrix at the node / ROI level for further analysis. Specifically, the present invention uses the functional connectivity matrix M of the atlas nThe node feature vector X of the nth atlas n , specifically for a particular atlas, the feature vectors of each node are its connection strengths / correlations with all nodes. For example, the functional connectivity matrix obtained for the nth atlas The features of its ith row represent the connection strength of the ith node with all R n nodes. In addition, the FCM generated by PCC is a complete graph, containing a lot of noise and irrelevant connections. The present invention retains a certain proportion (30%) of the strongest connections as edges to obtain a sparse binary adjacency matrix A. In this way, the multi-atlas graph information of each rs-fMRI can be structured as G n =(X n , A n ), where n ∈ {1, 2, 3} represents the index of each atlas.

[0024] 3. Input G1, G2, and G3 into the multi-atlas graph representation learning network as shown in Figure 2 . It uses a multi-layer graph convolutional network to extract high-level graph embedding features, and each layer of the network consists of "graph convolutional layer, ReLU activation layer, and dropout layer". Specifically, for each atlas information, the structured graph G=(X, A) is used as the input for graph feature representation learning. The node feature matrix of the (l + 1)th layer of the graph convolutional network can be calculated as follows:

[0025] H 0 =X

[0026] where is the adjacency matrix with self-loops added to include the features of each node itself; I R is the identity matrix; is 's degree matrix, obtained by filling the diagonal elements of A with ∑ i A ij and setting the non-diagonal elements to 0, and A ij represents the elements in A. While normalizing the feature contributions to ensure numerical stability, W l is the layer-specific trainable weight matrix, ReLU is the activation function, and Dt represents the dropout operation.

[0027] After extracting features through the multi-atlas graph network, the feature matrix of each atlas is normalized to a unified number of feature channels of 50. To obtain the graph-level feature representation of each atlas, X' n passes through a node-level global average pooling layer, which aggregates R nGenerate graph-level feature vectors from the features of all nodes

[0028] 4. Fuse the feature representations obtained at different graph scales (i.e., V1, V2, V3) using two different fusion strategies, MHCAF and AAWL, which specifically include the following sub-steps:

[0029] 41. As Figure 2 shown in the MHCAF module, for each graph feature vector three key components are generated, namely the query vector the key vector and the value vector These are all generated by three trainable projection matrices and Specifically, D q determines the dimension of the query feature space and is used to characterize the focus of the target feature on the context information. D k is used to represent the importance of the context feature and the degree of association with the target, while D v is used to store and transmit the context feature information actually used for fusion. Here, to simplify the model design and computational complexity, we uniformly set D q = D k = D v . Therefore, in each head (with H heads operating in parallel), we can fuse the target graph feature V j and the context graph feature V i to generate the output V ij as follows:

[0030]

[0031] where means not identically equal to.

[0032] Considering that in the fusion process, each graph feature is used as the target feature twice and the context feature twice, where V n as the context feature, the two calculation results are first combined to obtain V'. This integration ensures that the features obtained through cross-fusion are combined more smoothly:

[0033]

[0034] where m = (n % 3) + 1, o = (m % 3) + 1, n % 3 represents the remainder of n divided by 3, and m % 3 represents the remainder of m divided by 3. The final output F MHCAF can be expressed as:

[0035]

[0036] 42. As shown by the AAWL module in Figure 2 , first, all atlas features are concatenated to obtain V c , and then it is processed through the graph-level and node-level attention mechanisms. Specifically, first, the result of the graph attention mechanism is calculated to obtain V at , and then it is supplemented by using node feature attention to obtain the final output F AAWL , and the calculation method is as follows:

[0037]

[0038] Among them, σ represents the Sigmoid function; MLP represents a multi-layer perceptron network, and the full spelling of MLP is MultilayerPerceptron; ‘s-’ represents parameter sharing. ⊕ represents element-wise summation, represents concatenation in the channel dimension, represents element-wise product. MP and AP represent the max pooling and average pooling operations respectively, and the full spellings of MP and AP are Max Pooling and Average Pooling; the subscripts “at” and “nf” respectively refer to their applications in the graph and node feature dimensions.

[0039] 5. After obtaining the fusion features through two fusion schemes, they are concatenated and then used for depression prediction. After Linear linear transformation and SoftMax activation, the final prediction result Pred can be obtained, and the operation process can be expressed as follows:

[0040]

[0041] 6. After obtaining the output Pred through the entire model, the cross-entropy loss is used to calculate the error and backpropagate to update the network weights:

[0042]

[0043] Among them, N is the total number of subjects in the training process. Here, the subjects include the patients and healthy control groups collected in step 1, and y i represents the true label of the i-th subject, and p i ∈[0,1] represents the corresponding diagnostic prediction value.

[0044] The number of training iterations of the present invention is set to 100. When the iteration reaches 100 times, it stops, and a trained multi-atlas graph representation learning network is obtained; by inputting the three functional connectivity matrices corresponding to the rs-fMRI data to be analyzed into the trained multi-atlas graph representation learning network, the analysis result is obtained.

[0045] The technical effects of the present invention will be described below in conjunction with specific data:

[0046] A five-fold cross-validation experiment was conducted on the analysis method of the present invention based on the rs-fMRI data of recurrent depression collected by the Rest-meta-MDD consortium (including 189 depression patients and 426 healthy controls). The evaluation was carried out through five performance indicators: accuracy (ACC), recall, specificity (SPE), F1 score (F1), and area under the curve (AUC). The specific results are shown in Table 1:

[0047] Table 1 Comparison of performance indicators between the method of the present invention and the prior art

[0048] Method ACC(%) Recall(%) SPE(%) F1(％) AUC(%) G-L-A-Model 65.04±0.03 51.20±1.15 71.54±0.30 46.73±0.36 64.35±0.36 M-A-GNN 62.92±0.19 52.60±0.16 67.36±0.30 46.65±0.34 63.56±0.25 MDGL 65.20±0.06 56.25±0.62 69.31±0.13 49.58±0.21 66.07±0.05 The method of the present invention 68.45±0.20 58.42±1.23 72.62±0.74 52.86±0.41 68.47±0.38

[0049] Among them, G-L-A-Model, M-A-GNN, and MDGL are all methods for multi-atlas brain functional connectivity information fusion published in authoritative international journals. It can be found that our method can analyze the brain functional connectivity of depression more accurately and effectively than the current mainstream methods, achieving better performance than them.

[0050] For G-L-A-Model, specific reference can be made to: M. Liu, Q. Huang, L. Huang, S. Ren, L. Cui, H. Zhang, Y. Guan, Q. Guo, F. Xie, D. Shen, Dysfunctions of multiscaledynamic brain functionalnetworks in subjective cognitive decline, Brain Communications 6(1)(2024) fcae010.

[0051] For M-A-GNN, specific reference can be made to: D.-J. Lee, D.-H. Shin, Y.-H. Son, J.-W. Han, J.-H. Oh, D.-H. Kim, J.H. Jeong, T.-E. Kam, Spectral graph neural network-based multi-atlasbrain network fusion for major depressive disorder diagnosis, IEEE Journal ofBiomedical and Health Informatics 28(5)(2024) 2967 2978.

[0052] For MDGL, specific references can be found in: Y. Ma, Q. Wang, L. Cao, L. Li, C. Zhang, L. Qiao, M. Liu, Multiscaledynamic graph learning for brain disorder detection with func tionalmri, IEEE Transactions on Neural Systems and Rehabilitation Engineering 31(2023)3501-3512.

[0053] In summary, the BOLD signals derived from brain regions defined by different spatial scales or structural criteria in the method of the present invention can provide complementary topological information, thereby enhancing the feature expression of brain imaging data and improving the richness of features. Most of the current invention contents are based on single atlas information for functional connectivity data analysis, and the work based on multiple atlases only simply concatenates multiple features for downstream tasks, making it difficult to avoid problems such as inconsistent feature scales and incomplete fusion. The method proposed in the present invention is to synergistically combine the features of different atlases through more advanced strategies, so as to better reveal the complex connection patterns in the brains of patients with depression, provide support for the early diagnosis and personalized treatment of depression, and enhance the comprehensiveness and applicability of brain functional connectivity analysis methods. This not only helps to accurately identify diseases, but also provides important data support for future treatment strategies.

[0054] Those of ordinary skill in the art will realize that the embodiments described herein are to assist the reader in understanding the principles of the present invention and should be understood that the scope of protection of the present invention is not limited to such specific statements and embodiments. For those skilled in the art, various changes and modifications can be made to the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the scope of the claims of the present invention.

Claims

1. A method for analyzing brain functional connectivity data of depression based on multi-atlas fusion, characterized in that: include: S1. Collect the subjects' rs-fMRI data; As a training dataset; S2, preprocessing of rs-fMRI data; S3. For the preprocessed rs-fMRI data, three groups of BOLD signals of different scales were obtained according to the automatic anatomical labeling atlas, Harvard-Oxford atlas, and Craddock200 atlas. S4, for the three groups of BOLD signals of different scales obtained in step S3, the Pearson correlation algorithm is used to calculate and obtain the corresponding functional connection matrices; S5, representing the functional connectivity matrix as graph structure data; S6, using the graph structure data corresponding to each rs-fMRI data obtained in step S4 of the training data set as the input of the multi-graph spectrum graph representation learning network, using the patient or healthy control group corresponding to each rs-fMRI data in the training data set as the data label, and training the multi-graph spectrum graph representation learning network; S7. Input the three functional connection matrices corresponding to the rs-fMRI data to be analyzed into the trained multi-atlas representation learning network to obtain the analysis results.

2. According to claim 1, a method for analyzing brain functional connectivity data of depression based on multi-atlas fusion, characterized in that: The subjects in step S1 include patients and a healthy control group.

3. According to claim 2, a method for analyzing brain functional connectivity data of depression based on multi-atlas fusion, characterized in that: The three groups of BOLD signals of different scales in step S3 all use the connection strength between the current node and other nodes as the feature vector X of the current node.

4. The method for analyzing brain functional connectivity data of depression based on multi-atlas fusion according to claim 3, characterized in that: Step S5 is specifically as follows: retaining the strongest connections as edges in the function matrix obtained in step S4 according to a set ratio to obtain a sparse binary adjacency matrix A; and obtaining graph structure data according to the eigenvector X and the adjacency matrix A.

5. The method for analyzing brain functional connectivity data of depression based on multi-atlas fusion according to claim 4, characterized in that: The multi-graph representation learning network specifically includes two groups of identical structures, each of which includes a graph convolution layer, a ReLU activation layer, and a dropout layer.

6. The method for analyzing brain functional connectivity data of depression based on multi-atlas fusion according to claim 5, characterized in that: The multi-graph spectral graph representation learning network of step S6 also includes a node-level global average pooling layer: the feature matrix extracted by the multi-graph spectral graph representation learning network is passed through the node-level global average pooling layer to aggregate the feature matrices of all nodes to generate a graph-level feature vector.

7. The method for analyzing brain functional connectivity data of depression based on multi-atlas fusion according to claim 6, characterized in that: The multi-atlas graph representation learning network of step S6 also includes: an MHCAF module and an AAWL module; which are used to fuse graph-level feature vectors.

8. The method for analyzing brain functional connectivity data of depression based on multi-atlas fusion according to claim 7, characterized in that: The multi-atlas representation learning network of step S6 also includes: a prediction module; outputting a prediction result according to the fusion result of the MHCAF module and the AAWL module.

9. The method for analyzing brain functional connectivity data of depression based on multi-atlas fusion according to claim 8, characterized in that: The multi-graph spectrum representation learning network in step S6 uses cross entropy loss to calculate errors and back-propagate to update network weights.

10. The method for analyzing brain functional connectivity data of depression based on multi-atlas fusion according to claim 9, characterized in that: The cross entropy loss expression is: Among them, L represents the cross entropy loss, N represents the total number of subjects, and y i represents the true label of the i-th subject, p i Represents y i The corresponding predicted value.

Images