A brain age prediction method based on multi-modal fuzzy feature fusion

Abstract

The application provides a brain age prediction method based on multi-modal fuzzy feature fusion, and belongs to the cross field of medical imaging and computer science.The technical scheme comprises the following steps: S1: collecting the nuclear magnetic resonance imaging of a subject to form an original sample set; S2: pre-processing the original sample set; S3: extracting features of three different modes by using a deep convolutional neural network, and radially splicing the features; S4: designing a fuzzy fusion module based on Choquet integral; S5: designing a collaborative convolution fusion module; S6: obtaining a predicted age based on linear regression, and forming a complete brain age prediction neural network.The brain age prediction method has the beneficial effects that: the multi-modal features extracted by the fuzzy fusion overcome the limitations of a single mode, the uncertainty in the multi-modal fusion process, and the prediction ability of the brain age prediction model is improved, so that the complex brain age prediction task can be more effectively coped with.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of medical imaging and computer science, and in particular relates to a brain age prediction method based on multimodal fuzzy feature fusion. Background Technology

[0002] In the cutting-edge research field of brain age prediction, traditional single-modal analysis methods often fail to fully reflect the true state of the brain's physiological age due to the rich information and complexity brought about by multimodal neuroimaging data, such as structural magnetic resonance imaging (SMRI), functional magnetic resonance imaging (fMRI), magnetic resonance spectroscopy (MRI), and blood biomarkers. Traditional multimodal neuroimaging data fusion can be divided into: pixel-level simple weighted averaging, multi-band fusion, geometric transformation, and registration; feature-level feature extraction and fusion; and decision-level region or object fusion. Simple weighted averaging cannot fully consider the correlation between image spatial information and features, and may lose detailed information; multi-band fusion is very sensitive to noise, and it is difficult to select appropriate decomposition and reconstruction thresholds; region or object-based fusion has high requirements for the accuracy of lesion detection and may ignore important information in other non-target areas.

[0003] To overcome these challenges, researchers have begun exploring more advanced and sophisticated data fusion methods. These methods aim to combine information from different modalities to more accurately predict brain age and reveal the complex mechanisms of brain changes with age. They have constructed deep neural network models, such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), or generative adversarial networks (GANs), to automatically learn and extract deep features from multimodal data. Furthermore, to address the heterogeneity between multimodal data, researchers have proposed multimodal feature fusion strategies. These strategies include early fusion, mid-term fusion, and late fusion. Early fusion focuses on integrating multimodal information during the data input stage to consider more comprehensive information during feature extraction; mid-term fusion focuses on fusing features from different modalities after feature extraction but before decision-making to optimize the decision-making process; and late fusion integrates results after each modality makes independent decisions to ensure the accuracy of the final prediction. The choice of these fusion strategies requires a trade-off based on specific task requirements and data characteristics to achieve optimal brain age prediction results and reveal the complex mechanisms of brain changes with age. Summary of the Invention

[0004] The purpose of this invention is to provide a brain age prediction method based on multimodal fuzzy feature fusion. By measuring the importance of different features, fuzzy measures are applied to fuse them together, thereby obtaining a more comprehensive and accurate feature representation, which solves the technical problems existing in brain age prediction technology.

[0005] The inventive concept of this invention is as follows: This invention introduces a multimodal feature fusion method based on Choquet fuzzy integrals. Choquet fuzzy integrals, as an advanced fuzzy mathematics tool, can effectively handle nonlinearity and interactions between features, making it particularly suitable for multimodal data fusion scenarios. It characterizes the relative importance and dependencies between features by defining a fuzzy measure, rather than simply assigning fixed weights. In brain age prediction, key features extracted from different modalities are first ranked, and a fuzzy integral feature input matrix is ​​calculated based on the ranking information. Then, by multiplying this matrix with a pre-trained fuzzy measure matrix, deep fusion of multimodal features is achieved.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0007] S1: Collect brain MRI data from the subjects to form the original sample dataset, which includes information such as the subjects' brain MRI, age, and gender;

[0008] S2: The raw brain MRI images were preprocessed using the DPARSF 7.0 toolbox to generate three different modalities of image data: fALFF, ReHo, and T1w. These three modalities were then paired with their corresponding ages and genders to form a complete sample dataset. Finally, the preprocessed sample dataset was divided into training, validation, and test datasets.

[0009] S3: Read and preprocess the dataset, and match the data with corresponding labels to form input data for the neural network. Construct a deep convolutional neural network to extract features of the same dimension from data of different modalities, obtaining feature tensors Z for the three modalities. falff Z reho Z t1w And concatenate these three feature tensors radially into a single feature tensor Z. cat ;

[0010] S4: Construct a fuzzy fusion module based on Choquet integrals. First, initialize the fuzzy measure parameter matrix FM based on the relevance subset, and calculate the final fuzzy measure parameter matrix FM' using the formula. Next, iterate through each dimension of the fused features using nested loops, sort the features in each dimension in ascending order, generate an index matrix, construct a difference matrix based on the index, and finally form the feature input matrix D. Then, perform matrix multiplication between the feature input matrix D and the fuzzy measure matrix FM' to calculate the Choquet integral value CHI in each dimension. Finally, assign the calculated Choquet integral value to the original fused features to obtain the fuzzily fused feature Z. CHI ;

[0011] S5: Construct a collaborative convolutional fusion module. The collaborative convolutional module consists of a 3D convolutional layer with specific parameter configuration, a batch normalization layer, and an ELU activation function. The relevant parameters for the 3D convolutional layer are set as follows: kernel_size = (1,1,3), stride = 1, padding = 0, dilation = (1,1,2). [The last part, "Z," appears to be a typo and can be left as is.] CHI After processing by the collaborative convolution fusion module, the fuzzy collaborative feature Z is obtained. col Next, an adaptive average pooling operation is performed on the feature tensor to obtain a tensor Z with a shape and size of (B,C,1,1,1). age .

[0012] S6: Age prediction task based on linear regression. (Z...) age Dimensionality reduction is performed, and the result is transformed into a two-dimensional feature tensor of size (B,C) through a linear layer and an ELU activation function. This tensor is then input into a specific linear regression module to obtain the final prediction result y. pred The linear regression module includes a linear layer with input dimension C and output dimension 16, a linear layer with input dimension 16 and output dimension 1, and a ReLU activation function.

[0013] As a further optimization of the brain age prediction method based on multimodal fuzzy feature fusion provided by the present invention, the specific steps of step S1 are as follows:

[0014] Step S1.1: Determine the target population for the study, such as age range, health status, etc., and recruit and screen eligible subjects to participate in the study;

[0015] Step S1.2: Use MRI technology to collect brain data of the subjects, and collect the subjects' age and gender information at the same time;

[0016] Step S1.3: Integrate the collected brain MRI data, age and gender information into a raw dataset.

[0017] As a further optimization of the brain age prediction method based on multimodal fuzzy feature fusion provided by the present invention, the specific steps of step S2 are as follows:

[0018] Step S2.1: Use the DPARSF7.0 toolbox to open the acquired raw brain MRI data, and perform format conversion, slice timing, head motion correction, and spatial normalization;

[0019] Step S2.2: Calculate and generate fALFF image, ReHo image and T1w image;

[0020] Step S2.3: Pair the generated multimodal images with their corresponding age and gender information to ensure that each sample has a complete set of modal images and corresponding age and gender labels;

[0021] Step S2.4: Divide the preprocessed and paired sample dataset into three subsets: training dataset, validation dataset, and test dataset.

[0022] As a further optimization of the brain age prediction method based on multimodal fuzzy feature fusion provided by the present invention, the specific steps of step S3 are as follows:

[0023] Step S3.1: First, preprocess the input image, including loading .nii data, standardizing the read data, shape cropping, and data dimensionality upscaling. Then, process the feature maps (fALFF, ReHo, T1w) of the three modalities through a convolutional layer and an ELU activation function to ensure that the dimensions of the output feature maps remain consistent.

[0024] Step S3.2: Capture features at different scales through asymmetric convolutional layers. After fusing the output features from multiple scales, apply batch normalization layers, ELU activation functions, and compressed activation blocks to enhance the representational power of the features. Each asymmetric convolutional block consists of four sub-layers, each including a convolutional layer and a batch normalization layer; the compressed activation block consists of a global average pooling layer, a fully connected layer, ReLU, and a sigmoid activation function.

[0025] Step S3.3: Perform matrix multiplication between the features extracted by the two adaptive convolutional layers and the weight matrix obtained by the compressed activation block, then downsample through a max pooling layer to obtain the enhanced features, and finally concatenate them with the features obtained by the second max pooling downsampling to obtain the final extracted feature Z. x ;

[0026] Step S3.3: The three modalities are respectively processed by the above convolutional neural network to extract features, resulting in the feature tensors Z of the three modalities. falff Z reho Z t1w Their shapes and sizes are (B, C, W, H, D), where B represents the number of batch samples, C represents the number of input feature channels, and W, H, and D represent the width, height, and depth of the image, respectively.

[0027] Step S3.4: Concatenate and fuse these three feature tensors in the spatial dimension (dim = D) to obtain the feature tensor Z. cat Its shape and size are (B,C,W,H,3D).

[0028] As a further optimization of the brain age prediction method based on multimodal fuzzy feature fusion provided by the present invention, the specific steps of step S4 are as follows:

[0029] Step S4.1: Initialize the fuzzy measure parameter matrix FM. First, the input feature dimension N is given by the expected tensor dimension initially extracted. in and output dimension N out Then, based on the input feature dimension N in Calculate the number of relevant subsets Next, create a shape with a size of (N) sub N out The initial values ​​are all 1.0 / N. in The fuzzy measure parameter tensor FM;

[0030] Step S4.2: Calculate the fuzzy measure parameter matrix FM' according to the formula, which is as follows:

[0031]

[0032] Where i is the index of the fuzzy measure parameter, s[i] is a list storing subset indices, and L(x) is a function that converts subset indices into an index list; the detailed process is as follows: First, the absolute value of the fuzzy measure parameter tensor FM is processed. Then, according to the feature subset index i, the corresponding fuzzy measure parameter subsets are taken out in sequence. If it is a single-element subset, the fuzzy measure parameter is directly added to the fuzzy measure parameter matrix FM'; if it is a multi-element subset, the maximum parameter value in the subset is found, and it is added to the current fuzzy measure parameter, and the result is added to the corresponding position in the fuzzy measure parameter matrix FM'.

[0033] Step S4.3: Add a column of all 1s to the end of the processed FM'. By comparing each element and taking the minimum value, FM' is restricted to ensure that all elements do not exceed 1.

[0034] Step S4.4: Feature Map Processing and Sorting. Process each position of the feature map one by one, i.e., sort the feature Z according to the batch size, number of channels, and height dimension. cat The feature tensor values ​​at each position are traversed and sorted to obtain the sorted feature matrix Feature and its corresponding index matrix Sorted.

[0035] Step S4.5: Construct the difference matrix, output index matrix, and feature input matrix. Calculate the difference of the sorted feature value sequence using the following formula to form the difference matrix ΔC;

[0036] ΔC=B ij -A ij (5)

[0037] Among them, A ij and B ij Let represent the eigenvalues ​​of the matrix at the i-th row and j-th column, respectively. Simultaneously, calculate the cumulative sum of the sorted indices, subtract the unit vector, and generate the output index matrix Sorted' used to construct the Choquet integral feature input matrix; then, initialize a matrix N with the same number of columns as the number of feature samples and equal to the number of columns in the fuzzy measure matrix. sub The zero matrix D with +1 is used to fill the corresponding positions of the feature input matrix D with the values ​​in the difference matrix ΔC according to the values ​​in the output index matrix Sorted'.

[0038] Step S4.6: Calculate the Choquet integral. Perform matrix multiplication on the constructed feature matrix D and the fuzzy measure matrix FM' to obtain the fused feature value CHI;

[0039]

[0040] Among them, CHI g (F) represents the output value after Choquet fuzzy integral feature fusion, where F represents the feature matrix, with each row representing a sample and each column representing a feature. μ(k) A represents the fuzzy measure of subset k, which measures the importance of features. μ(k) s(A) represents the feature map after the subset k is fused. μ(0) )=0,s(A μ(k) )-s(A μ(k-1) ) represents the difference between adjacent subsets, μ is a permutation that requires g to be such that μ(1) ≥g μ(2) ≥...≥g μ(N) ;

[0041] Step S4.7: Obtain the fuzzy fusion feature map. Reassign the calculated fuzzy fusion feature value CHI to the corresponding position in the original feature map to obtain the fuzzy fusion feature Z. CHI .

[0042] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0043] 1. High Efficiency of Multimodal Feature Fuzzy Fusion: By employing a multimodal feature fusion method based on Choquet fuzzy integrals, this invention overcomes the limitations of traditional linear fusion methods, deeply exploring and revealing the profound correlations between features of different modalities. This fusion method not only improves information utilization efficiency but also reduces the subjectivity and uncertainty caused by manually setting weights to a certain extent, providing a more accurate and comprehensive feature foundation for subsequent brain age prediction.

[0044] 2. Significantly improved accuracy of brain age prediction: By introducing fuzzy fusion technology, this invention effectively handles the fuzziness in multimodal data, balances the prediction results of the network, effectively prevents overfitting, and significantly improves the accuracy and stability of brain age prediction.

[0045] 3. Enhanced Model Flexibility and Scalability: When faced with new imaging technologies or data types, this method can quickly adapt to and integrate new information sources without requiring large-scale modifications to the overall framework. Furthermore, due to its modular design, each part can be independently optimized and updated, further enhancing the model's scalability and providing strong support for research and applications in brain age prediction. Attached Figure Description

[0046] 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.

[0047] Figure 1 This is a flowchart of a brain age prediction method based on multimodal fuzzy feature fusion according to the present invention;

[0048] Figure 2 This is a diagram illustrating the overall network framework of a brain age prediction method based on multimodal fuzzy feature fusion according to the present invention.

[0049] Figure 3 This is a diagram of a deep convolutional neural network structure for a brain age prediction method based on multimodal fuzzy feature fusion according to the present invention. Detailed Implementation

[0050] 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. Of course, the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0051] Example 1

[0052] like Figure 1 As shown, the brain age prediction method based on multimodal fuzzy feature fusion in this embodiment includes the following steps:

[0053] S1: Collect brain MRI data from the subjects to form the original sample dataset, which includes information such as the subjects' brain MRI, age, and gender;

[0054] S2: The raw brain MRI images were preprocessed using the DPARSF 7.0 toolbox to generate image data in three different modalities (fALFF, ReHo, T1w). Then, these three modalities were paired with their corresponding ages and genders to form a complete sample dataset. Finally, the preprocessed sample dataset was divided into training, validation, and test datasets.

[0055] S3: Read the dataset and preprocess it, then match the data with corresponding labels to form input data for the neural network. Construct a deep convolutional neural network, such as... Figure 3 As shown, features of the same dimension are extracted from data of different modalities to obtain the feature tensors Z of the three modalities. falff Z reho Z t1w The shape and size are (8, 1944, 2, 3, 2), and these three feature tensors are radially concatenated into a single feature tensor Z. cat The shape and size are (8,1944,2,3,6);

[0056] S4: Construct a fuzzy fusion module based on Choquet integrals. First, initialize the fuzzy measure parameter matrix FM based on the relevance subset, and calculate the final fuzzy measure parameter matrix FM' using the formula. Next, iterate through each dimension of the fused features using nested loops, sort the features in each dimension in ascending order, generate an index matrix, construct a difference matrix based on the index, and finally form the feature input matrix D. Then, perform matrix multiplication between the feature input matrix D and the fuzzy measure matrix FM' to calculate the Choquet integral value CHI in each dimension. Finally, assign the calculated Choquet integral value to the original fused features to obtain the fuzzily fused feature Z. CHI ;

[0057] S5: Construct a collaborative convolutional fusion module. The collaborative convolutional module consists of a 3D convolutional layer with specific parameter configuration, a batch normalization layer, and an ELU activation function. The relevant parameters for the 3D convolutional layer are set as follows: kernel_size = (1,1,3), stride = 1, padding = 0, dilation = (1,1,2). [The last part, "Z," appears to be a typo and can be left as is.] CHI After processing by the collaborative convolution fusion module, the fuzzy collaborative feature Z is obtained. col The shape and size become (8,1944,2,3,2). Then, an adaptive average pooling operation is performed on this feature tensor to obtain a tensor Z with a shape and size of (8,1944,1,1,1). age .

[0058] S6: Age prediction task based on linear regression. (Z...) ageThe dimension is reduced to a shape of (8, 1944), and then passed through a linear layer and an ELU activation function to transform it into a two-dimensional feature tensor with a shape of (8, 32). This tensor is then input into a specific linear regression module to obtain the final prediction result y. pred The linear regression module includes a linear layer with an input dimension of 32 and an output dimension of 16, a linear layer with an input dimension of 16 and an output dimension of 1, and a ReLU activation function.

[0059] Specifically, the steps of step S1 are as follows:

[0060] Step S1.1: Determine the target population for the study, such as the age range of 18-80 years old, both healthy and unhealthy, and recruit and screen eligible subjects to participate in the study;

[0061] Step S1.2: Use MRI technology to collect brain data of the subjects, and collect the subjects' age and gender information at the same time;

[0062] Step S1.3: Integrate the collected brain MRI data, age and gender information into a raw dataset.

[0063] Specifically, the steps of step S2 are as follows:

[0064] Step S2.1: Use the DPARSF7.0 toolbox to open the acquired raw brain MRI data, perform format conversion, slice timing, head motion correction and spatial normalization. The slice timing is set to 2s, the head motion correction removes data with motion exceeding 2.5mm, and all pixel values ​​of the image are normalized to [0,1].

[0065] Step S2.2: Calculate and generate fALFF image, ReHo image and T1w image;

[0066] Step S2.3: Pair the generated multimodal images with their corresponding age and gender information to ensure that each sample has a complete set of modal images and corresponding age and gender labels;

[0067] Step S2.4: Divide the preprocessed and paired sample dataset into three subsets: training dataset, validation dataset, and test dataset, where 75% is used for training, 15% for validation, and 10% for testing.

[0068] Specifically, the steps of step S3 are as follows:

[0069] Step S3.1: First, preprocess the input image, including loading .nii data, standardizing the read data, shape cropping, and data dimensionality upscaling. Then, process the feature maps (fALFF, ReHo, T1w) of the three modalities through a convolutional layer and an ELU activation function to ensure that the dimensions of the output feature maps remain consistent. The shape and size of the preprocessed feature tensor are (8, 1, 81, 99, 81).

[0070] Step S3.2: Capture features at different scales through asymmetric convolutional layers. After fusing the output features from multiple scales, apply batch normalization layers, ELU activation functions, and compressed activation blocks to enhance the representational power of the features. Each asymmetric convolutional block consists of four sub-layers, each including a convolutional layer and a batch normalization layer; the compressed activation block consists of a global average pooling layer, a fully connected layer, ReLU, and a sigmoid activation function.

[0071] Step S3.3: Perform matrix multiplication between the features extracted by the two adaptive convolutional layers and the weight matrix obtained by the compressed activation block, then downsample through a max pooling layer to obtain the enhanced features, and finally concatenate them with the features obtained by the second max pooling downsampling to obtain the final extracted feature Z. x The change in the size of the feature tensor is as follows:

[0072] (8,8,81,99,81)→(8,24,40,49,40)→(8,72,20,24,20)→(8,216,10,12,10)→(8,648,5,6,5)→(8,1944,2,3,2);

[0073] Step S3.3: The three modalities are respectively processed by the above convolutional neural network to extract features, resulting in the feature tensors Z of the three modalities. falff Z reho Z t1w Their shapes and sizes are all (8, 1944, 2, 3, 2);

[0074] Step S3.4: Concatenate and fuse these three feature tensors in the spatial dimension (dim = D) to obtain the feature tensor Z. cat Its shape and size are (8,1944,2,3,6).

[0075] Specifically, the steps of step S4 are as follows:

[0076] Step S4.1: Initialize the fuzzy measure parameter matrix FM. First, the input feature dimension N is given by the expected tensor dimension initially extracted. in =6 and output dimension N out =6, then based on the input feature dimension N inCalculate the number of relevant subsets Next, create a shape with a size of (62, 6) and an initial value of 1.0 / N. in The fuzzy measure parameter tensor FM;

[0077]

[0078] Step S4.2: Calculate the fuzzy measure parameter matrix FM' according to the formula, which is as follows:

[0079]

[0080] Where i is the index of the fuzzy measure parameter, s[i] is a list storing subset indices, and L(x) is a function that converts subset indices into an index list. The detailed process is as follows: First, the absolute value of the fuzzy measure parameter tensor FM is processed. Then, according to the feature subset index i, the corresponding fuzzy measure parameter subsets are retrieved sequentially. If it is a single-element subset, the fuzzy measure parameter is directly added to the fuzzy measure parameter matrix FM'; if it is a multi-element subset, the maximum parameter value within the subset is calculated, and it is added to the current fuzzy measure parameter, with the result added to the corresponding position in the fuzzy measure parameter matrix FM'.

[0081] Step S4.3: Add a column of all 1s to the end of the processed FM'. By comparing each element and taking the minimum value, FM' is restricted to ensure that all elements do not exceed 1.

[0082]

[0083] Step S4.4: Feature Map Processing and Sorting. Process each position of the feature map one by one, i.e., sort the feature Z according to the batch size, number of channels, and height dimension. cat The feature tensor values ​​at each position are traversed and sorted to obtain the sorted feature matrix Feature and its corresponding index matrix Sorted. Taking the input feature F as an example, the corresponding Feature and Sorted are obtained.

[0084]

[0085]

[0086]

[0087] Step S4.5: Construct the difference matrix, output index matrix, and feature input matrix. Calculate the difference of the sorted feature value sequence using the following formula to form the difference matrix ΔC;

[0088] ΔC=B ij -A ij (13)

[0089]

[0090] Among them, A ij and B ij Let represent the eigenvalues ​​of the matrix at the i-th row and j-th column, respectively. Simultaneously, calculate the cumulative sum of the sorted indices, subtract the unit vector, and generate the output index matrix Sorted' used to construct the Choquet integral feature input matrix; then, initialize a matrix N with the same number of columns as the number of feature samples and equal to the number of columns in the fuzzy measure matrix. sub The zero matrix D with +1 is used to fill the corresponding positions of the feature input matrix D with the values ​​in the difference matrix ΔC according to the values ​​in the output index matrix Sorted'.

[0091]

[0092]

[0093] Step S4.6: Calculate the Choquet integral. Perform matrix multiplication on the constructed feature matrix D and the fuzzy measure matrix FM' to obtain the fused feature value CHI;

[0094]

[0095] Among them, CHI g (F) represents the output value after Choquet fuzzy integral feature fusion, where F represents the feature matrix, with each row representing a sample and each column representing a feature. μ(k) A represents the fuzzy measure of subset k, which measures the importance of features. μ(k) s(A) represents the feature map after the subset k is fused. μ(0) )=0,s(A μ(k) )-s(A μ(k-1) ) represents the difference between adjacent subsets, μ is a permutation that requires g to be such that μ(1) ≥g μ(2) ≥...≥g μ(N) ;

[0096]

[0097] Step S4.7: Obtain the fuzzy fusion feature map. Reassign the calculated fuzzy fusion feature value CHI to the corresponding position in the original feature map to obtain the fuzzy fusion feature Z. CHI Its tensor size is (8,1944,2,3,6).

[0098] In Example 2, the 1,138 participants were aged between 18 and 80 years (623 males and 515 females), including 636 healthy participants and 502 unhealthy participants. The training set, validation set, and test set were divided into 75%, 15%, and 10% groups, respectively, as detailed in Table 1.

[0099] Number 853 171 114 1138

[0100] Table 1

[0101] During model training, the batch size was set to 8, the initial learning rate to 0.001, the weight decay parameter to 0.0005, and the number of iterations to 150. The model evaluation metric was MAE, calculated using the following formula:

[0102]

[0103] Where N is the number of subject samples; It is the subject's predicted age; y i This refers to the actual age of the subject. As shown in Table 2, the prediction method of this invention, after multimodal fuzzy fusion, has better accuracy than the single-modal prediction.

[0104]

[0105]

[0106] Table 2

[0107] 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 prediction method based on multimodal fuzzy feature fusion, characterized in that, Includes the following steps: S1: Collect brain MRI data from the subjects to form the original sample dataset, which includes the subjects' brain MRI data, age, and gender information; S2: The raw brain MRI images were preprocessed using the DPARSF7.0 toolbox to generate three different modalities of image data: fALFF, ReHo, and T1w. Then, the images of these three modalities were paired with their corresponding ages and genders to form a complete sample dataset. Finally, the preprocessed sample dataset was divided into training dataset, validation dataset, and test dataset. S3: Read and preprocess the dataset, and match the data with corresponding labels to form input data for inputting into the neural network. Construct a deep convolutional neural network to extract features of the same dimension from data of different modalities, obtaining feature tensors of three modalities. , , And concatenate these three feature tensors radially into a single feature tensor. ; S4: Construct a fuzzy fusion module based on Choquet integrals. First, initialize the fuzzy measure parameter matrix according to the correlation subset. The final fuzzy measure parameter matrix is ​​calculated using the formula. Next, nested loops are used to traverse and fuse the feature tensors. The features are sorted in ascending order according to each dimension to generate an index matrix. A difference matrix is ​​then constructed based on the indexes to form the final feature input matrix. Then, the feature input matrix is... With fuzzy measure matrix Perform matrix multiplication to calculate the Choquet integral values ​​in each dimension. Finally, the calculated Choquet integral value is assigned to the fusion feature tensor. The features obtained after fuzzy fusion are obtained. ; S5: Construct a collaborative convolutional fusion module. This module consists of a 3D convolutional layer with specific parameter configuration, a batch normalization layer, and an ELU activation function. The relevant parameters for the 3D convolutional layer are set as follows: kernel_size=(1,1,3), stride=1, padding=0, dilation=(1,1,2), will After processing by the collaborative convolutional fusion module, fuzzy collaborative features are obtained. Next, an adaptive average pooling operation is performed on the feature tensor to obtain a shape with a size of tensor ; S6: Age prediction task based on linear regression, Dimensionality reduction is performed, and the shape is transformed to a size of through a linear layer and an ELU activation function. The two-dimensional feature tensor is input into a specific linear regression module to obtain the final prediction result. The linear regression module includes a linear layer with an input dimension of C and an output dimension of 16, a linear layer with an input dimension of 16 and an output dimension of 1, and a ReLU activation function.

2. The brain age prediction method based on multimodal fuzzy feature fusion according to claim 1, characterized in that, The specific steps of step S1 are as follows: Step S1.1: Determine the target population for the study, such as age range and health status, and recruit and screen eligible subjects to participate in the study; Step S1.2: Use MRI technology to collect brain data of the subjects, and collect the subjects' age and gender information at the same time; Step S1.3: Integrate the collected brain MRI data, age and gender information into a raw dataset.

3. The brain age prediction method based on multimodal fuzzy feature fusion according to claim 1, characterized in that, The specific steps of step S2 are as follows: Step S2.1: Use the DPARSF 7.0 toolbox to open the acquired raw brain MRI data, and perform format conversion, slice timing, head motion correction, and spatial normalization; Step S2.2: Calculate and generate fALFF image, ReHo image and T1w image; Step S2.3: Pair the generated multimodal images with their corresponding age and gender information to ensure that each sample has a complete set of modal images and corresponding age and gender labels; Step S2.4: Divide the preprocessed and paired sample dataset into three subsets: training dataset, validation dataset, and test dataset.

4. The brain age prediction method based on multimodal fuzzy feature fusion according to claim 1, characterized in that, The specific steps of step S3 are as follows: Step S3.1: First, preprocess the input image, including loading .nii data, standardizing the read data, cropping the shape, and increasing the data dimensionality; then, process the feature maps fALFF, ReHo, and T1w of the three modalities through a convolutional layer and an ELU activation function to ensure that the dimensions of the output feature maps remain consistent. Step S3.2: Capture features at different scales through asymmetric convolutional layers, add and fuse the output features at multiple scales, and then apply batch normalization layers, ELU activation functions, and compressed activation blocks to enhance the representational power of the features. Each asymmetric convolutional block consists of four sub-layer structures, each of which includes a convolutional layer and a batch normalization layer; the compressed activation block consists of a global average pooling layer, a fully connected layer, ReLU, and a sigmoid activation function. Step S3.3: Perform matrix multiplication between the features extracted by the two adaptive convolutional layers and the weight matrix obtained by the compressed activation block, then downsample through a max pooling layer to obtain enhanced features, and finally concatenate with the features obtained by the second max pooling downsampling to obtain the final extracted features. ; Step S3.3: Feature extraction is performed on the three modalities respectively through the above convolutional neural network to obtain the feature tensors of the three modalities. Their shape and size are all Where B represents the number of batch samples, C represents the number of input feature channels, and W, H, and D represent the width, height, and depth of the image, respectively; Step S3.4: Concatenate and fuse these three feature tensors along the spatial dimension dim=D to obtain the feature tensor. Its shape and size are .

5. The brain age prediction method based on multimodal fuzzy feature fusion according to claim 1, characterized in that, The specific steps of step S4 are as follows: Step S4.1: Initialize the fuzzy measure parameter matrix First, the input feature dimension is given by the tensor dimension initially extracted as expected. and output dimensions Then, based on the input feature dimensions Calculate the number of relevant subsets Next, create a shape with a size of The initial values ​​are all Fuzzy measure parameter tensor ; Step S4.2: Calculate the fuzzy measure parameter matrix according to the formula. The calculation formula is as follows: （1）； in, It is an index of the fuzzy measure parameter. It is a list that stores subset indices. It is a function that converts subset indices into a list of indices; first, it converts the fuzzy measure parameter tensor... Perform absolute value processing, and then index according to feature subset. Extract the corresponding subsets of fuzzy measure parameters one by one. If it is a single-element subset, directly add the fuzzy measure parameters to the fuzzy measure parameter matrix. If it is a multi-element subset, find the maximum parameter value within the subset, add it to the current fuzzy measure parameter, and add the result to the fuzzy measure parameter matrix. The corresponding position in the middle; Step S4.3: After processing Add a column of all 1s to the end, and then compare each element to find the minimum value. Implement restrictions to ensure that all elements do not exceed 1; Step S4.4: Feature map processing and sorting. Process each position of the feature map one by one, that is, according to the batch size, channel size, and height dimension, sort the features. The feature tensor values ​​at each position are traversed and sorted to obtain the sorted feature value matrix. and its corresponding index matrix ; Step S4.5: Construct the difference matrix, output index matrix, and feature input matrix. Calculate the difference of the sorted feature value sequence using the following formula to form the difference matrix. ; （2）； in, and These represent the matrix at the th order. line, number The column's eigenvalues ​​are calculated, and the cumulative sum of the sorted indices is subtracted to generate the output index matrix used to construct the Choquet integral feature input matrix. Then, initialize a matrix with the same number of feature samples and a number of columns equal to the number of columns in the fuzzy measure matrix. All-zero matrix According to the output index matrix The values ​​in the output will be the difference matrix. The values ​​in the matrix are filled into the feature input matrix. The corresponding position; Step S4.6: Calculate the Choquet integral using the constructed feature matrix. With fuzzy measure matrix Perform matrix multiplication to obtain the fused eigenvalues. ; （3）； in, This represents the output value after fusion based on Choquet fuzzy integral features. This represents a feature matrix, where each row represents a sample and each column represents a feature. Representing a subset The fuzzy measure value measures the importance of features. Representing a subset The fused feature map , Represents the difference between adjacent subsets. It is a permutation that needs to make ; Step S4.7: Obtain the fuzzy fusion feature map and calculate the fuzzy fusion feature values. The values ​​are then reassigned to the corresponding positions in the original feature map to obtain the fuzzy fused features. .

Images