A brain age estimation method based on a dynamic fuzzy learnable brain network

Abstract

The application provides a brain age estimation method based on a dynamic fuzzy learnable brain network, and belongs to the technical field of medical image processing and artificial intelligence. The technical scheme comprises the following steps: S1: collecting the brain magnetic resonance imaging of a subject, and performing pretreatment and data division; S2: constructing graph structure data, and performing feature extraction and position information embedding on the data; S3: constructing a dynamic fuzzy learnable brain network model comprising a main branch and a local branch, and respectively extracting global and local connection features; S4: introducing a dynamic fuzzy multi-head self-attention module into the main branch, and realizing effective modeling of the global features; S5: dynamically modeling the inter-channel dependency relationship through a convolution filter and a learnable graph attention module in the local branch; and S6: after fusing the features of the main branch and the local branch, performing brain age prediction through a multilayer perception machine. The application improves the modeling capability of the brain functional connection mode, and can more effectively complete the brain age prediction task.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of medical image processing and artificial intelligence, and specifically relates to a brain age estimation method based on dynamic fuzzy learnable brain networks. Background Technology

[0002] With the increase in average human lifespan and the accelerating development of an aging society, scientifically assessing brain health and identifying neurodegenerative diseases at an early stage have become important research topics in neuroscience and clinical medicine. Brain age gap, as an indicator reflecting the degree of physiological aging of the brain, is widely used in the auxiliary diagnosis and efficacy evaluation of various neurological diseases such as Alzheimer's and Parkinson's. By comparing the difference between predicted brain age and actual age, abnormal changes in individual brain structure and function can be revealed, providing a scientific basis for disease monitoring, treatment intervention, and personalized aging management. Therefore, constructing accurate and robust brain age estimation models is of great significance for improving the clinical diagnosis and treatment of neurological diseases.

[0003] Functional magnetic resonance imaging (fMRI), as an important non-invasive method for acquiring information on brain activity, plays a crucial role in brain age modeling research because it can reveal the dynamic functional connectivity between different brain regions. The functional connectivity matrix constructed based on fMRI can capture complex interaction patterns between brain regions, revealing the dynamic characteristics of the brain's spatiotemporal structure. However, this type of data is typically characterized by high feature dimensionality, significant signal redundancy, and substantial individual differences. Furthermore, the brain's connectivity structure is highly heterogeneous due to the combined influence of genetic and environmental factors. These issues greatly increase the complexity of modeling and limit the generalization and expressive power of traditional statistical methods and shallow machine learning models in brain age prediction tasks.

[0004] Existing brain age prediction methods suffer from insufficient information fusion, incomplete feature extraction, and low prediction accuracy when processing structurally and functionally complex neuroimaging data. Traditional linear methods such as linear regression, support vector machines, and principal component analysis struggle to handle nonlinear structures and multimodal heterogeneous data. In recent years, deep learning models such as CNNs, Transformers, and graph neural networks (GNNs) have been increasingly applied to this field, but challenges remain in multi-scale fusion, long-term dependency modeling, and local pattern extraction. Summary of the Invention

[0005] The purpose of this invention is to provide a brain age estimation method based on dynamic fuzzy learnable brain networks, addressing the technical problems in existing technologies such as the difficulty in co-modeling global and local features of brain functional connectivity data, insufficient noise robustness, and poor adaptive learning ability of local key regions. Through a collaborative architecture combining a dynamic fuzzy fusion mechanism with learnable graph attention, efficient fusion of multi-scale brain features can be achieved, providing more reliable age biomarkers for the early detection of neurodegenerative diseases.

[0006] To achieve the above-mentioned objectives, the present invention employs the following technical solution: a brain age estimation method based on a dynamic fuzzy learnable brain network, comprising the following steps:

[0007] S1: Collect brain MRI images of the subjects, preprocess the data, including: slice timing correction, normalization and noise reduction, and divide the dataset into training set, validation set and test set in a ratio of 8:1:1;

[0008] S2: Construct graph-structured data, and perform feature extraction and location information embedding on the data;

[0009] S3: Construct a dynamic fuzzy learnable brain network model containing main branches and local branches, and extract global and local connectivity features respectively;

[0010] S4: The main branch introduces a dynamic fuzzy multi-head self-attention module, which integrates Gaussian, triangular and trapezoidal fuzzy membership functions to achieve feature weighting;

[0011] S5: Local branches dynamically model inter-channel dependencies using convolutional filters and learnable graph attention modules;

[0012] S6: After fusing features from the main branch and local branches, brain age is predicted using a multilayer perceptron.

[0013] As a further optimization of the brain age estimation method based on dynamic fuzzy learnable brain networks provided by the present invention, step S2 includes the following steps:

[0014] Step S2.1: Based on the Automated Anatomical Label Atlas (AAL), the DPARSF tool was used to extract time series data of 116 brain regions and calculate the correlation between regions to construct a functional connectivity matrix.

[0015] Step S2.2: Construct the brain region functional connectivity data into a graph structure, where nodes represent brain regions and edge weights represent connectivity strength;

[0016] Step S2.3: Input graph data, node features The edge connection weights E are input to the TransformerConv layer to extract local connectivity features, which are then reshaped into a 3D tensor.

[0017]

[0018] Where F represents the feature dimension, N represents the number of input nodes, B is the batch size, and H is the hidden feature dimension.

[0019] As a further optimization of the brain age estimation method based on dynamic fuzzy learnable brain networks provided by the present invention, step S3 includes the following steps:

[0020] Step S3.1: Input the preprocessed features into both the main branch and the local branches simultaneously;

[0021] Step S3.2: The main branch adopts an identity-based location information embedding strategy. By concatenating the preprocessed features along the feature dimension, an enhanced global feature representation is obtained:

[0022]

[0023] Step S3.3: Local branches are upgraded through dimensionality increase operations. Local patterns are extracted using convolutional filters, which include E2E (Eege-to-Edge) convolutional filters, E2N (Edge-to-Node) convolutional filters, and N2G (Node-to-Graph) convolutional filters. After each E2E convolutional operation, the learnable graph attention module LGAM is used to adaptively enhance local features, preserve key node connections, and suppress redundant information.

[0024] As a further optimization of the brain age estimation method based on dynamic fuzzy learnable brain networks provided by the present invention, step S4 includes the following steps:

[0025] Step S4.1: Input feature tensor Generate the query Q, key K, and value V matrix through linear transformation:

[0026] Q = X global W Q K = X global W K V = X global W V (3)

[0027] Step S4.2: Model global dependencies using a multi-head self-attention mechanism:

[0028]

[0029] Step S4.3: Utilize Gaussian, trigonometric, and trapezoidal membership functions to achieve dynamic feature weighting, where the Gaussian membership function μ... GThe expression for smoothing the central region of the distribution is:

[0030]

[0031] Trigonometric membership function Used to emphasize boundary information, its expression is:

[0032]

[0033] Trapezoidal membership function Used to capture regional transition features, its expression is:

[0034]

[0035] in, σ represents the center of the Gaussian and trigonometric membership functions. g Controlling the density of the Gaussian distribution, Adjust the span of the trigonometric functions. Four key points representing the membership function of a trapezoid;

[0036] Step S4.4: Perform linear processing on the blurred features:

[0037]

[0038] in, The weight matrix corresponding to the linear transformation, For the corresponding bias term;

[0039] Step S4.5: The fused features are weighted through a learnable linear layer, and then normalized using softmax and activated by the tanh activation function to enhance their nonlinear expressive power.

[0040]

[0041] Where α, β, and γ are the importance weights for each fuzzy feature channel.

[0042] As a further optimization of the brain age estimation method based on dynamic fuzzy learnable brain networks provided by the present invention, step S5 includes the following steps:

[0043] Step S5.1: Input local feature tensor First, global average pooling is used to aggregate the average features of each channel along the spatial dimension. Then, a 1×1 one-dimensional convolution is used to further compress the channel information to obtain the feature representation Y', whose expression is:

[0044] Y' = Conv1d(AvgPool(X) local (10)

[0045] Step S5.2: Initialize the static adjacency matrix as an identity matrix, the expression is: By introducing a convolution operation with learnable parameters, a dynamic adjacency matrix is ​​computed, the expression of which is:

[0046] A dynamic =softmax(Conv2d(Y')) (11)

[0047] Step S5.3: Combine the static adjacency matrix, the dynamic adjacency matrix, and a trainable matrix. Merge into the final adjacency matrix:

[0048] A = A static ⊙A dynamic +A trainable (12)

[0049] Where ⊙ represents the Hadamard product, i.e., element-wise multiplication;

[0050] Step S5.4: Using the adjacency matrix obtained above to guide information interaction between channels, calculate the attention mask M:

[0051] M=σ(Conv1d(ReLU(Y'·A))) (13)

[0052] Where · represents matrix multiplication, ReLU activation introduces a nonlinear transformation, and σ represents the Sigmoid function, which is used to restrict the attention value to the interval [0,1].

[0053] Step S5.5: Apply the generated attention mask to the original local feature tensor to achieve dynamic weighting:

[0054] Z local =X local ⊙M (14).

[0055] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0056] 1. Dual-branch structure for collaborative modeling of global and local features: This invention designs a dual-branch framework based on Transformer and graph convolution structures. The main branch captures global dependency patterns of brain functional connectivity through a multi-head self-attention mechanism, while the local branches extract local connectivity relationships through graph attention mechanisms and multi-level convolution operations, thereby achieving joint modeling of macroscopic and microscopic features and effectively improving the accuracy of brain age prediction.

[0057] 2. Enhanced Feature Representation Through Dynamic Fuzzy Fusion: This invention introduces a dynamic fuzzy fusion mechanism based on Gaussian, triangular, and trapezoidal membership functions, which can effectively model fuzzy boundaries and nonlinear transition regions in neuroimages. Compared to traditional fixed-weight or linear weighting methods, this mechanism has stronger feature adaptability and nonlinear interactive modeling capabilities, significantly improving the accuracy and robustness of feature representation.

[0058] 3. Significantly Enhanced Model Adaptability and Flexibility: This invention is based on a modular architecture design, decoupling the main branch, local branches, and feature fusion module, enabling each component to be independently replaceable and flexibly upgraded. This design not only improves the model's adaptability to different neuroimaging modalities but also facilitates targeted optimization of each module according to specific task requirements, thereby significantly improving the overall system's generalization performance and operational efficiency across multiple scenarios. Attached Figure Description

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used to explain the invention but do not constitute a limitation thereof.

[0060] Figure 1 This is a flowchart of the present invention;

[0061] Figure 2 This is a schematic diagram of the overall model framework of the present invention;

[0062] Figure 3 This is a diagram of the dynamic fuzzy multi-head self-attention module of the present invention;

[0063] Figure 4 This is a diagram of the learnable graph attention module of the present invention;

[0064] Figure 5 This is a scatter plot of the present invention;

[0065] Figure 6 This is a residual distribution diagram for the present invention. Detailed Implementation

[0066] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. The specific embodiments described herein are merely illustrative and are not intended to limit the scope of the invention.

[0067] Example 1

[0068] The flowchart of this embodiment is as follows: Figure 1 The model framework diagram is as follows Figure 2 Dynamically blurred multi-head self-attention module, such as Figure 3 Learnable graph attention modules, such as Figure 4This embodiment presents a brain age estimation method based on a dynamic fuzzy learnable brain network, comprising the following steps:

[0069] S1: Collect brain MRI images of the subjects, preprocess the data, including: slice timing correction, normalization and noise reduction, and divide the dataset into training set, validation set and test set in a ratio of 8:1:1;

[0070] S2: Construct graph-structured data, and perform feature extraction and location information embedding on the data;

[0071] S3: Construct a dynamic fuzzy learnable brain network model containing main branches and local branches, and extract global and local connectivity features respectively;

[0072] S4: The main branch introduces a dynamic fuzzy multi-head self-attention module, which integrates Gaussian, triangular and trapezoidal fuzzy membership functions to achieve feature weighting;

[0073] S5: Local branches dynamically model inter-channel dependencies using convolutional filters and learnable graph attention modules;

[0074] S6: After fusing features from the main branch and local branches, brain age is predicted using a multilayer perceptron.

[0075] Specifically, the steps of step S2 are as follows:

[0076] Step S2.1: Based on the Automated Anatomical Label Atlas (AAL), the DPARSF tool was used to extract time series data of 116 brain regions and calculate the correlation between regions to construct a functional connectivity matrix.

[0077] Step S2.2: Construct the brain region functional connectivity data into a graph structure, where nodes represent brain regions and edge weights represent connectivity strength;

[0078] Step S2.3: Input graph data, node features The edge connection weights E are input to the TransformerConv layer to extract local connectivity features, which are then reshaped into a 3D tensor.

[0079]

[0080] Where F represents the feature dimension, N represents the number of input nodes, B is the batch size, and H is the hidden feature dimension.

[0081] Specifically, the steps of step S3 are as follows:

[0082] Step S3.1: Input the preprocessed features into both the main branch and the local branches simultaneously;

[0083] Step S3.2: The main branch adopts an identity-based location information embedding strategy. By concatenating the preprocessed features along the feature dimension, an enhanced global feature representation is obtained:

[0084]

[0085] Step S3.3: Local branches are upgraded through dimensionality increase operations. Local patterns are extracted using convolutional filters, which include E2E (Eege-to-Edge) convolutional filters, E2N (Edge-to-Node) convolutional filters, and N2G (Node-to-Graph) convolutional filters. After each E2E convolutional operation, the learnable graph attention module LGAM is used to adaptively enhance local features, preserve key node connections, and suppress redundant information.

[0086] Specifically, the steps of step S4 are as follows:

[0087] Step S4.1: Input feature tensor Generate the query Q, key K, and value V matrix through linear transformation:

[0088] Q = X global W Q K = X global W K V = X global W V (3)

[0089] Step S4.2: Model global dependencies using a multi-head self-attention mechanism:

[0090]

[0091] Step S4.3: Utilize Gaussian, trigonometric, and trapezoidal membership functions to achieve dynamic feature weighting, where the Gaussian membership function μ... G The expression for smoothing the central region of the distribution is:

[0092]

[0093] Trigonometric membership function Used to emphasize boundary information, its expression is:

[0094]

[0095] Trapezoidal membership function Used to capture regional transition features, its expression is:

[0096]

[0097] in, σ represents the center of the Gaussian and trigonometric membership functions. g Controlling the density of the Gaussian distribution, Adjust the span of the trigonometric functions. Four key points representing the membership function of a trapezoid;

[0098] Step S4.4: Perform linear processing on the blurred features:

[0099]

[0100] in, The weight matrix corresponding to the linear transformation, For the corresponding bias term;

[0101] Step S4.5: The fused features are weighted through a learnable linear layer, and then normalized using softmax and activated by the tanh activation function to enhance their nonlinear expressive power.

[0102]

[0103] Where α, β, and γ are the importance weights for each fuzzy feature channel.

[0104] Specifically, the steps of step S5 are as follows:

[0105] Step S5.1: Input local feature tensor First, global average pooling is used to aggregate the average features of each channel along the spatial dimension. Then, a 1×1 one-dimensional convolution is used to further compress the channel information to obtain the feature representation Y', whose expression is:

[0106] Y' = Conv1d(AvgPool(X) local (10)

[0107] Step S5.2: Initialize the static adjacency matrix as an identity matrix, the expression is: By introducing a convolution operation with learnable parameters, a dynamic adjacency matrix is ​​computed, the expression of which is:

[0108] A dynamic =softmax(Conv2d(Y')) (11)

[0109] Step S5.3: Combine the static adjacency matrix, the dynamic adjacency matrix, and a trainable matrix. Merge into the final adjacency matrix:

[0110] A = A static ⊙A dynamic +A trainable (12)

[0111] Where ⊙ represents the Hadamard product, i.e., element-wise multiplication;

[0112] Step S5.4: Using the adjacency matrix obtained above to guide information interaction between channels, calculate the attention mask M:

[0113] M=σ(Conv1d(ReLU(Y'·A))) (13)

[0114] Where · represents matrix multiplication, ReLU activation introduces a nonlinear transformation, and σ represents the Sigmoid function, which is used to restrict the attention value to the interval [0,1].

[0115] Step S5.5: Apply the generated attention mask to the original local feature tensor to achieve dynamic weighting:

[0116] Z local =X local ⊙M (14).

[0117] Example 2

[0118] This embodiment uses the SRPBS multiple impairment MRI dataset, which contains brain imaging data of 1410 participants in NIFTI format. Among them, there are 790 healthy participants (428 men and 362 women) and 620 unhealthy participants (382 men and 238 women).

[0119] The brain imaging fMRI data in NIFTI format were preprocessed. The preprocessing steps included slice timing correction, normalization, and denoising to eliminate unacceptable data. The demographic data of the processed SRPBS dataset are shown in Table 1.

[0120] Table 1

[0121]

[0122] During model training, the following hyperparameters were set: batch size of 128, number of training epochs of 100, and learning rate of Adam optimizer of 0.001. In the Transformer encoder, the dimension of the feedforward network was set to 2048, the dropout parameter was set to 0.1, and the epsilon value for layer normalization was 1e-5.

[0123] During model validation and testing, the model evaluation metric is the mean absolute error (MAE), which is calculated using the following formula:

[0124]

[0125] Where N is the number of subject samples, y iIt is the actual age of the subject. It is the age predicted by the test subjects.

[0126] The evaluation metric is the standard deviation (STD), which is calculated using the following formula:

[0127]

[0128] in, This represents the average prediction error.

[0129] The evaluation index is the Pearson correlation coefficient (PCC), which is calculated using the following formula:

[0130]

[0131] in, This represents the average actual age. This represents the predicted average age.

[0132] Table 2 shows a comparative experiment between the model in this embodiment and BC-GCN and its variants, BrainNetCNN and BrainNetTransformer. The prediction method proposed in this embodiment is superior to all models.

[0133] Table 2

[0134]

[0135] Example 3

[0136] On the dataset of Example 2, ablation experiments were conducted on the model proposed in this example to verify the performance contributions of the learnable attention modules (LGAM1, LGAM2) and the dynamic fuzzy fusion layer (DFFL). The performance results are shown in Table 3, and scatter plots and residual plots of the model proposed in this example are also plotted as follows: Figure 5 and Figure 6 As shown.

[0137] Table 3

[0138]

[0139] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A brain age estimation method based on dynamic fuzzy learnable brain networks, characterized in that, Includes the following steps: S1: Collect brain MRI images of the subjects, preprocess the data, including: slice timing correction, normalization and noise reduction, and divide the dataset into training set, validation set and test set in a ratio of 8:1:1; S2: Construct graph-structured data, and perform feature extraction and location information embedding on the data; S3: Construct a dynamic fuzzy learnable brain network model containing main branches and local branches, and extract global and local connectivity features respectively; S4: The main branch introduces a dynamic fuzzy multi-head self-attention module, which achieves dynamic weighting of global features by fusing Gaussian, triangular and trapezoidal fuzzy membership functions; Step S4 includes the following steps: Step S4.1: Input feature tensor Generate the query Q, key K, and value V matrix through linear transformation: (3)； Step S4.2: Model global dependencies using a multi-head self-attention mechanism: (4)； Step S4.3: Utilize Gaussian, trigonometric, and trapezoidal membership functions to achieve dynamic feature weighting. Specifically, the Gaussian membership function... The expression for smoothing the central region of the distribution is: (5)； Trigonometric membership function Used to emphasize boundary information, its expression is: (6)； Trapezoidal membership function Used to capture regional transition features, its expression is: (7)； in, Indicate the center of Gaussian and trigonometric membership functions. Controlling the density of the Gaussian distribution, Adjust the span of the trigonometric functions. Four key points representing the membership function of a trapezoid; Step S4.4: Perform linear processing on the blurred features: ， ， (8)； in, The weight matrix corresponding to the linear transformation, For the corresponding bias term; Step S4.5: The fused features are weighted through a learnable linear layer, and then normalized using softmax and activated by the tanh activation function to enhance their nonlinear expressive power. (9)； in, Importance weights for each fuzzy feature channel; S5: Local branches dynamically model inter-channel dependencies using convolutional filters and learnable graph attention modules; S6: After fusing features from the main branch and local branches, brain age is predicted using a multilayer perceptron.

2. The brain age estimation method based on dynamic fuzzy learnable brain networks according to claim 1, characterized in that, Step S2 includes the following steps: Step S2.1: Based on the Automated Anatomical Label Atlas (AAL), the DPARSF tool was used to extract time series data of 116 brain regions and calculate the correlation between regions to construct a functional connectivity matrix. Step S2.2: Construct the brain region functional connectivity data into a graph structure, where nodes represent brain regions and edge weights represent connectivity strength; Step S2.3: Input graph data, node features Edge connection weight The input to the TransformerConv layer extracts local connectivity features, which are then reshaped into a 3D tensor. (1)； in, Represents feature dimension, Represents the number of input nodes. For batch size, To hide the feature dimension.

3. The brain age estimation method based on dynamic fuzzy learnable brain networks according to claim 1, characterized in that, Step S3 includes the following steps: Step S3.1: Input the preprocessed features into both the main branch and the local branches simultaneously; Step S3.2: The main branch adopts an identity-based location information embedding strategy. The enhanced global feature representation is obtained by concatenating the preprocessed features along the feature dimension. (2)； Step S3.3: Local branches are upgraded through dimensionality increase operations. Local patterns are extracted using convolutional filters, which include E2E, E2N, and N2G convolutional filters. After each E2E convolution operation, the learnable graph attention module LGAM is used to adaptively enhance local features, preserve key node connections, and suppress redundant information.

4. The brain age estimation method based on dynamic fuzzy learnable brain networks according to claim 1, characterized in that, Step S5 includes the following steps: Step S5.1: Input local feature tensor First, a global average pooling operation is used to converge the average features of each channel along the spatial dimension. Then, a 1×1 one-dimensional convolution is used to further compress the channel information to obtain the feature representation. Its expression is: (10)； Step S5.2: Initialize the static adjacency matrix as an identity matrix, the expression is: By introducing a convolution operation with learnable parameters, a dynamic adjacency matrix is ​​computed, the expression of which is: (11)； Step S5.3: Combine the static adjacency matrix, the dynamic adjacency matrix, and a trainable matrix. These are then merged into the final adjacency matrix: (12)； in, This represents the Hadamard product, i.e., element-wise multiplication; Step S5.4: Using the adjacency matrix obtained above to guide information interaction between channels, calculate the attention mask M: (13)； in, ReLU activation, representing matrix multiplication, introduces a nonlinear transformation. This represents the Sigmoid function, used to restrict attention values ​​to the [0,1] interval; Step S5.5: Apply the generated attention mask to the original local feature tensor to achieve dynamic weighting: (14)。

Images