Construction method and application of multi-scale brain age prediction model based on magnetic resonance imaging

Abstract

The method and application of constructing a multi-scale brain age prediction model based on magnetic resonance imaging (MRI) have enhanced generalization and robustness, maintained high prediction accuracy, and can predict brain age at the whole brain, subnetwork, and voxel levels, making the predicted brain age more interpretable in a physiological sense. The method explores the differences in brain age between different subnetworks and the specific patterns of their PAD and cognitive associations, finds specific subnetworks that regulate cognition, and observes the differential patterns of aging in different brain regions at the voxel level. The model's prediction performance is superior to the current mainstream neural network model. The method includes: (1) data collection; (2) data preprocessing; (3) constructing a whole brain and functional subnetwork brain age prediction model based on the simple fully convolutional neural network (SFCN) method; (4) constructing a voxel-level brain age prediction model based on the ScaledDense U-Net method; and (5) bias correction for brain age deviation.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing and machine learning applications, and in particular to a method for constructing a multi-scale brain age prediction model based on magnetic resonance images, and the application of the method for constructing a multi-scale brain age prediction model based on magnetic resonance images. Background Art

[0002] Initial brain age prediction models were built based on traditional machine learning algorithms, primarily using gray matter and white matter information extracted from brain images as features. Commonly used machine learning models include Support Vector Regression (SVR), Gaussian Process Regression (GPR), Ridge Regression, Elastic Net, Relevance Vector Machine (RVM), and LASSO regression. Brain age prediction models based on these traditional machine learning algorithms have achieved promising results in brain age prediction research, providing an effective approach for exploring scientific questions about brain age. Among them, Franke et al. used the T1w images of 410 healthy subjects collected from different sites in the IXI dataset as a training set. After dimensionality reduction using the principal component analysis method, they used the RVM model to train a brain age prediction model. They used 245 healthy subjects as a test set to predict their brain age and applied the model to 334 subjects in the ADNI dataset (including 102 AD patients and 232 normal people). The results showed that the model had good stability and cross-site generalization ability. The Pearson correlation coefficient between the predicted brain age and the actual age in normal people was 0.92, the mean absolute error (MAE) was 5 years, and the average difference between the brain age and the actual age (BrainAGE) in the AD population reached 10. It was also found that the training sample size was an important factor in the accuracy of the imaging model. Cole et al. extracted GM and WM as input features from T1w images of 1,537 healthy subjects, constructed a brain age prediction model based on GPR, and applied it to patients with traumatic brain injury (TBI). The model showed a Pearson correlation coefficient of 0.92 between the predicted brain age and the actual age in healthy subjects. Compared with normal subjects, the brains of TBI patients were older, with the difference (PAD) between the predicted age and the actual age in WM and GM being 5.97 and 4.66, respectively, indicating that the brain is aging at an accelerated rate throughout the late stage of chronic injury. Huang et al. used elastic networks to train a brain age prediction model with gray matter features from 974 healthy subjects (490 of whom were from the BABRI dataset and 484 from the ADNI dataset), and predicted the brain age and PAD of 231 healthy subjects and 224 patients with aMCI. The results showed that compared with the normal control group, aMCI patients had a higher PAD. When the AD risk-related allele APOE was taken into account, the PAD of patients with aMCI was higher. 4, the group carrying the gene showed a higher PAD than the group not carrying the gene. It was also found that the PAD of amyloid-positive aMCI patients was higher than that of amyloid-negative patients.

[0003] Over the past few years, machine learning algorithms have driven the development of brain age prediction and achieved significant results, demonstrating their unique advantages. However, the application of traditional machine learning algorithms in brain age prediction research requires manual feature extraction from brain images. Only after meaningful features are extracted can subsequent training be performed. In recent years, with the development of deep learning, various deep learning algorithm models have been applied to brain age prediction research. Compared to traditional machine learning algorithms, deep learning algorithms have complex models that can capture higher-level feature expressions from raw data and can accommodate larger data sets. Therefore, deep learning algorithms have attracted great attention from researchers and have driven significant progress in brain age prediction research. Jonsson et al. extracted Jacobian images, gray matter images, and white matter images from T1w brain images of 1,264 healthy people from Iceland, and used a 3D convolutional neural network (CNN) to train the model. The four predicted brain ages were linearly combined to obtain the final brain age. In this study, the CNN model introduced a short-circuit mechanism to alleviate the gradient vanishing phenomenon caused by network deepening, and used data augmentation technology to prevent model overfitting. The trained brain age prediction model was tested on the UK Biobank dataset and transfer learning was performed on a new dataset, which improved the model generalization and prediction accuracy. The prediction result showed an MAE of 3.5 years. After performing a whole-genome association analysis on the predicted PAD, two mutation sites were found, one of which was related to reduced sulcus width and the other was related to reduced white matter surface area. To overcome the large sample size and computational resource requirements of deep learning, Peng et al. proposed a Simple Fully Convolutional Network (SFCN). Compared to other mainstream deep learning networks, this model has fewer parameters and is more suitable for large datasets. Using raw T1w images as input, they trained a brain age model on 14,503 subjects from the UK Biobank, achieving a mean average error (MAE) of 2.14 years and a gender classification accuracy of 99.5%. Most brain age prediction studies have focused on middle-aged and elderly individuals. To investigate brain age across all age groups, Bashyam et al. used the DeepBrainNet algorithm to train a brain age prediction model on a dataset of 11,729 subjects ranging in age from 3 to 95 years from various datasets. They found that a moderately fit brain age prediction model captured more significant differences across disease groups than an overfit model. They also demonstrated that transfer learning of the DeepBrainNet-based brain age prediction model can construct more accurate brain disease classifiers.Most current brain age prediction studies construct brain age prediction models based on the whole brain or regions of interest (ROI). Although some models of brain age studies generate images of areas sensitive to brain age prediction and discover some physiological and anatomical significance from the distribution range of voxels that contribute most to brain age prediction, the physiological age of each voxel in the brain cannot be estimated. Therefore, Popescu et al. used the U-Net algorithm model and T1w images as input features to train a brain age prediction model on a dataset of 3463 healthy people and generated individual 3D brain age maps. They found that the median mean absolute error of individual brain age was 9.5 years. After correcting for PAD at the voxel level, they found that there were significantly different local brain age patterns between healthy controls and patients with mild cognitive impairment or dementia, especially in subcortical regions such as the nucleus accumbens, putamen, globus pallidus, hippocampus, and amygdala. Nguyen et al. used an integrated 3DU-Net network to predict brain age at the voxel level on T1w brain images, obtained the brain age of the brain's anatomical structure through 3D segmentation masks, and calculated the deviations in the normal aging process of different brain structures. This model can be used to accurately classify brain diseases at the individual level.

[0004] However, the sample size used for modeling in current brain age prediction research is relatively limited. Generally, a single brain age model is constructed based on the whole brain, and only an average brain age value for the whole brain is output. The brain age estimates provided by the brain age prediction model have poor physiological significance and application value, and the prediction accuracy is not high. Summary of the Invention

[0005] In order to overcome the defects of the existing technology, the technical problem to be solved by the present invention is to provide a method for constructing a multi-scale brain age prediction model based on magnetic resonance images. The constructed brain age prediction model has stronger generalization and robustness, maintains a high prediction accuracy, and can predict the brain age of the whole brain-subnetwork-voxel, so that the predicted brain age has better interpretability in a physiological sense, explores the differences in brain age between different subnetworks and the specific patterns of their PAD and cognitive association, finds the specific subnetworks that regulate cognition, and observes the differential patterns of aging in different brain regions at the voxel level. The model's prediction performance is better than the current mainstream neural network model.

[0006] The technical solution of the present invention is: a method for constructing a multi-scale brain age prediction model based on magnetic resonance images, which comprises the following steps:

[0007] (1) Data collection: The T1-weighted MRI dataset of healthy subjects was used as the training set, and the mixed T1-weighted MRI dataset was used as the test set. The mixed T1-weighted MRI dataset contained T1-weighted MRI images of normal subjects, patients with subjective cognitive decline, mild cognitive impairment, and Alzheimer's disease.

[0008] (2) Data preprocessing: The original images were preprocessed using Freesurfer software, including head motion correction, non-uniform field intensity normalization, linear registration, and scalping. The images were then resampled to a voxel space with a resolution of 2 mm and a size of 128 × 128 × 128 using the QSIPrep software package. The Yeo functional partitioning template was then nonlinearly registered to the individual space and uniformly resampled to a resolution of 2 mm to extract seven functional subnetworks: the visual network, the somatomotor network, the dorsal attention network, the ventral attention network, the limbic network, the frontoparietal network, and the default mode network.

[0009] (3) Constructing a whole-brain and functional subnetwork brain age prediction model based on the simple fully convolutional neural network (SFCN) method: The whole-brain T1-weighted magnetic resonance images of each subject in the training set with a resolution of 2 mm and the seven functional subnetwork images obtained by segmentation based on the Yeo template were used as the input of the 3D neural network. The input image size was 128 × 128 × 128. A 3D deep convolutional neural network model was constructed based on the SFCN method. A series of 3D convolutional layers were used to extract high-dimensional features from the image. Finally, a fully connected layer was used for regression to output the predicted brain age.

[0010] (4) Construct a voxel-level brain age prediction model based on the ScaledDense U-Net method: A whole-brain T1-weighted magnetic resonance image with a size of 91×109×91 and a resolution of 2mm aligned to the MNI standard space is used as the input of the ScaledDenseU-Net method model. The output of the model is a T1-weighted magnetic resonance image with the same size and resolution as the input image. The model presents a U-shaped structure, with the left part being the encoder part and the right part being the decoder part. The feature maps of the same size in the left and right symmetrical parts of the U-shaped structure are connected by jump connections. During the upsampling process, the underlying features are fused with the high-level features of the corresponding positions, thereby retaining more high-resolution information. The model first performs a three-dimensional convolution operation on the input image with a stride of 1, a convolution kernel size of 7×7×7, and 8 channels, and then activates it with ELU to obtain a feature map with a size of 81×99×81 and 8 channels, which is then passed through 4 dense layer blocks and maximum pooling. The decoder gradually restores the spatial resolution by performing deconvolution and upsampling operations on the feature maps. After each upsampling, the decoder feature map is spliced ​​with the high-level features of the same size corresponding to the encoder on the left. The spliced ​​tensor passes through the upper part of the dense layer. This operation is repeated four times and the output feature map is upsampled to the same size as the original input. The predicted brain age map is output, where the kernel size of the deconvolution is 2×2×2 and the size of the output feature map is twice that of the input.

[0011] (5) Bias correction of brain age deviation: The linear regression method is used to correct the bias of brain age deviation PAD. The bias correction method relies on the linear regression model of PAD and actual age. The formula for PAD bias correction is as follows:

[0012] (1)

[0013] in Represents the real age, is the slope, is the intercept, It is PAD, with linear regression model fitting and Afterwards, the corrected brain age is obtained by subtracting this bias from the predicted brain age, using the following formula:

[0014] (2)

[0015] in represents the corrected brain age, represents the predicted brain age, and is the slope and intercept of the linear model fitted on the training set. Formula (2) is applied to the test set to obtain the corrected brain age on the test set, and then the corrected brain age is subtracted from the predicted brain age to obtain the corrected PAD.

[0016] The beneficial technical effects of the present invention are as follows:

[0017] 1) Build a brain age prediction model on a large dataset to improve its generalizability. Current brain age prediction studies have used insufficiently large sample sizes. This study built a brain age prediction model based on approximately 28,000 individuals from the UK Biobank dataset. This model demonstrates enhanced generalizability and robustness, enabling it to maintain high prediction accuracy when applied to other independent datasets.

[0018] 2) A multi-scale brain age prediction model was constructed based on the whole brain, functional subnetworks, and voxels. Compared with whole-brain-based models, this study's multi-scale model can predict brain age at the "whole brain-subnetwork-voxel" level, making the predicted brain age more physiologically interpretable.

[0019] 3) Build brain age prediction models based on seven sub-networks respectively, explore the differences in brain age between different sub-networks and the specific patterns of their PAD-cognition associations, and find specific sub-networks that regulate cognition.

[0020] 4) A voxel-level brain age prediction model was constructed based on the ScaledDense U-Net, which examines the differential aging patterns across different brain regions at the voxel level. The ScaledDense U-Net introduced deep supervision and a ranking loss function, resulting in a model with better prediction performance than current mainstream neural network models.

[0021] It also provides the application of a method for constructing a multi-scale brain age prediction model based on magnetic resonance images. Its application in Alzheimer's disease risk prediction uses PAD as a biomarker for assessing brain health during brain aging. There are significant differences in PAD at the voxel level between normal people and people with mild cognitive impairment, normal people and people with Alzheimer's disease, and people with mild cognitive impairment and people with Alzheimer's disease. The areas with significant differences are distributed in brain regions other than the middle frontal gyrus, the superior temporal gyrus, and the orbital gyrus. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 4 is a flowchart of a method for constructing a multi-scale brain age prediction model based on magnetic resonance images according to the present invention. DETAILED DESCRIPTION

[0023] like Figure 1 As shown in FIG, the method for constructing a multi-scale brain age prediction model based on magnetic resonance images includes the following steps:

[0024] (1) Data collection: The T1-weighted MRI dataset of healthy subjects was used as the training set, and the mixed T1-weighted MRI dataset was used as the test set. The mixed T1-weighted MRI dataset contained T1-weighted MRI images of normal subjects, patients with subjective cognitive decline, mild cognitive impairment, and Alzheimer's disease.

[0025] (2) Data preprocessing: The original images were preprocessed using Freesurfer software, including head motion correction, non-uniform field intensity normalization, linear registration, and scalping. The images were then resampled to a voxel space with a resolution of 2 mm and a size of 128 × 128 × 128 using the QSIPrep software package. The Yeo functional partitioning template was then nonlinearly registered to the individual space and uniformly resampled to a resolution of 2 mm to extract seven functional subnetworks: the visual network, the somatomotor network, the dorsal attention network, the ventral attention network, the limbic network, the frontoparietal network, and the default mode network.

[0026] (3) Constructing a whole-brain and functional subnetwork brain age prediction model based on the simple fully convolutional neural network (SFCN) method: The whole-brain T1-weighted magnetic resonance images of each subject in the training set with a resolution of 2 mm and the seven functional subnetwork images obtained by segmentation based on the Yeo template were used as the input of the 3D neural network. The input image size was 128 × 128 × 128. A 3D deep convolutional neural network model was constructed based on the SFCN method. A series of 3D convolutional layers were used to extract high-dimensional features from the image. Finally, a fully connected layer was used for regression to output the predicted brain age.

[0027] (4) Construct a voxel-level brain age prediction model based on the ScaledDense U-Net method: A whole-brain T1-weighted magnetic resonance image with a size of 91×109×91 and a resolution of 2mm aligned to the MNI standard space is used as the input of the ScaledDenseU-Net method model. The output of the model is a T1-weighted magnetic resonance image with the same size and resolution as the input image. The model presents a U-shaped structure, with the left part being the encoder part and the right part being the decoder part. The feature maps of the same size in the left and right symmetrical parts of the U-shaped structure are connected by jump connections. During the upsampling process, the underlying features are fused with the high-level features of the corresponding positions, thereby retaining more high-resolution information. The model first performs a three-dimensional convolution operation on the input image with a stride of 1, a convolution kernel size of 7×7×7, and 8 channels, and then activates it with ELU to obtain a feature map with a size of 81×99×81 and 8 channels, which is then passed through 4 dense layer blocks and maximum pooling. The decoder gradually restores the spatial resolution by performing deconvolution and upsampling operations on the feature maps. After each upsampling, the decoder feature map is spliced ​​with the high-level features of the same size corresponding to the encoder on the left. The spliced ​​tensor passes through the upper part of the dense layer. This operation is repeated four times and the output feature map is upsampled to the same size as the original input. The predicted brain age map is output, where the kernel size of the deconvolution is 2×2×2 and the size of the output feature map is twice that of the input.

[0028] (5) Bias correction of brain age deviation: The linear regression method is used to correct the bias of brain age deviation PAD. The bias correction method relies on the linear regression model of PAD and actual age. The formula for PAD bias correction is as follows:

[0029] (1)

[0030] in Represents the real age, is the slope, is the intercept, It is PAD, with linear regression model fitting and Afterwards, the corrected brain age is obtained by subtracting this bias from the predicted brain age, using the following formula:

[0031] (2)

[0032] in represents the corrected brain age, represents the predicted brain age, and is the slope and intercept of the linear model fitted on the training set. Formula (2) is applied to the test set to obtain the corrected brain age on the test set, and then the corrected brain age is subtracted from the predicted brain age to obtain the corrected PAD.

[0033] The beneficial technical effects of the present invention are as follows:

[0034] 1) Build a brain age prediction model on a large dataset to improve its generalizability. Current brain age prediction studies have used insufficiently large sample sizes. This study built a brain age prediction model based on approximately 28,000 individuals from the UK Biobank dataset. This model demonstrates enhanced generalizability and robustness, enabling it to maintain high prediction accuracy when applied to other independent datasets.

[0035] 2) A multi-scale brain age prediction model was constructed based on the whole brain, functional subnetworks, and voxels. Compared with whole-brain-based models, this study's multi-scale model can predict brain age at the "whole brain-subnetwork-voxel" level, making the predicted brain age more physiologically interpretable.

[0036] 3) Build brain age prediction models based on seven sub-networks respectively, explore the differences in brain age between different sub-networks and the specific patterns of their PAD-cognition associations, and find specific sub-networks that regulate cognition.

[0037] 4) A voxel-level brain age prediction model was constructed based on the ScaledDense U-Net, which examines the differential aging patterns across different brain regions at the voxel level. The ScaledDense U-Net introduced deep supervision and a ranking loss function, resulting in a model with better prediction performance than current mainstream neural network models.

[0038] Preferably, in step (1), the UK Biobank dataset is used as a training set to construct a brain age prediction model, and the Xuanwu Hospital dataset, ADNI and BABRI datasets are used as independent datasets to evaluate the generalization ability of the brain age prediction model and cross-validate the PAD group differences between different cognitive groups.

[0039] Preferably, in step (2), the T1-weighted magnetic resonance image that has been registered to the standard space and resampled to a resolution of 2 mm is subjected to Z-Score normalization, and the normalization formula is as follows:

[0040] (3)

[0041] in, This is the image before standardization. It's an image The mean of all voxel grayscale values, is the standard deviation of the grayscale values ​​of all voxels in image y, The definition is as follows:

[0042] (4)

[0043] in, is the value of the corresponding voxel in the brain mask, Indicates the Grayscale value of the voxel.

[0044] Preferably, in step (3), the first five modules are composed of a 3×3×3 convolution layer, a batch normalization layer, an ELU activation function layer and a maximum pooling layer, the sixth block is composed of a 1×1×1 convolution layer, a batch normalization layer, an ELU activation function layer and an average pooling layer, the pooling kernel size of the maximum pooling layer and the average pooling layer is 2×2×2, and the first six blocks are used for feature extraction; the seventh block is composed of two fully connected layers and a Dropout layer, the first fully connected layer outputs a tensor of length 16, which is then passed through the ELU activation function and the Dropout layer. The Dropout layer retains or discards a certain proportion of neurons according to a given probability, and the output value of the discarded neurons is set to 0 during the forward propagation process. The last fully connected layer outputs the final predicted brain age.

[0045] Preferably, in step (3), the model initializes the network model parameters with the_uniform, which is a normal distribution initializer that starts with a center of 0 and a standard deviation of The images are sampled from a truncated normal distribution, fan_in is the number of input neurons in the weight tensor; the dropout ratio of the Dropout layer is set to 0.2, that is, neurons are randomly dropped with a probability of 20% to avoid overfitting; the batch size of the network model is set to 32, and 32 image samples are loaded in each round of training; the optimizer is the Adam optimizer, the initial learning rate is 0.01, and the weight decay coefficient is 1Í10 -6 The loss function is the L1 loss function. If the loss function does not decrease after 120 consecutive iterations during the training process, the learning rate is automatically multiplied by a factor of 0.1 to obtain a smaller learning rate, and the iterative training continues. The maximum number of iterations is set to 500.

[0046] The T1-weighted magnetic resonance images of the UK Biobank dataset were divided into training, validation, and test sets according to a 6:2:2 ratio. To reduce the impact of the model's prediction accuracy on the training, validation, and test sets due to the uneven age distribution of the subjects in the dataset, the dataset was randomly divided according to a ratio to ensure that the age distribution of the subjects in the training, validation, and test sets was roughly the same. The whole-brain brain age prediction model and the brain age prediction model of the seven functional subnetworks were trained separately.

[0047] Preferably, in step (4), each layer of densely connected blocks consists of two asymmetric convolution AC blocks, two batch normalization layers and an ELU activation function layer, and a compression-excitation SE module. The AC block uses four three-dimensional convolution layers. The features output by the SE module are only spliced ​​with the features before the first AC block. The number of channels of the input feature tensor is n, and the number of channels of the final output feature tensor is 3n. The number of output feature channels of the upsampling part of the densely connected block is half of the number of input feature channels. The number of channels of the input feature tensor is halved after passing through a convolution layer. The output of the convolution layer is spliced ​​with the feature tensor output by the SE network block, and after splicing, it continues to pass through a convolution layer to obtain the final output feature tensor.

[0048] Preferably, in step (4), for N samples, the Spearman rank correlation coefficient SRCC is defined as the Spearman rank correlation coefficient between the ranked values ​​of the two variables:

[0049] (5)

[0050] in Rank(·) is a sort operation, is the covariance of the two sets of ranked values, is the standard deviation of the two sets of sorted values, if and There are no identical values ​​in , the above definition is equivalent to:

[0051] (6)

[0052] If there are identical values ​​in two groups of variables, their ranking value is the mean of the ranks of these identical values ​​in the ascending sequence;

[0053] The ranking loss function is defined as follows:

[0054] (7)

[0055] The SoDeep model is used to replace the Rank(·) operation. SoDeep trains a differentiable proxy network R based on artificially synthesized data. The network R is designed based on the LSTM network. After replacing the Rank(·) operation with the network R, the ranking loss function is defined as follows:

[0056] (8)

[0057] where R(·) is an approximate representation of the variable ordering, and SoDeep Reach the minimum to train the network R until the model converges;

[0058] The mean square error (MSE) loss function is calculated for the feature maps output by the 1st, 2nd, and 3rd layers of the decoder in the ScaledDense U-Net network, and the loss function is obtained. and as well as , the MSE loss function between the brain age map output by the model and the target value is , the ranking loss function is , the final loss function It is the linear weighted sum of these four loss functions, and the formula is as follows:

[0059] (9)

[0060] in 、 、 、 and is the weight coefficient before each loss function.

[0061] Preferably, in step (4), a voxel-level smoothing process is performed on the brain age map outputted last in the network, the smoothing kernel size is 3×3×3, the loss function in the training process is the MSE loss between the smoothed image and the target image, and the smoothing operation does not participate in the gradient update in the training process. The model uses the Adam optimizer, the batch size is set to 16, the initial learning rate is set to 0.0001, and the weight decay is 5×10 -4 The learning rate update strategy is to multiply the learning rate by 0.5 to obtain a new learning rate when the loss function on the validation set does not decrease after 10 epochs. Continue training the model until 30 epochs have passed and the loss function on the validation set still cannot be reduced. Then stop training the model.

[0062] Preferably, the method further comprises step (6), model evaluation, wherein the evaluation indicators include the Pearson correlation coefficient, determination coefficient and mean absolute error (MAE) between the predicted brain age and the actual age;

[0063] The Pearson correlation coefficient is defined as follows:

[0064] (10)

[0065] The coefficient of determination reflects the proportion of the variation of the dependent variable explained by the independent variable through the regression relationship and is defined as follows:

[0066] (11)

[0067] The mean absolute error MAE is as follows:

[0068] (12)

[0069] in is the brain age map predicted by the model, It's a mask for the brain. It is the average brain age at the voxel level of the whole brain. This average brain age value will be used as the predicted brain age value in the subsequent statistical analysis based on the individual whole brain.

[0070] The results of the present invention show that the brain age prediction model of the whole brain and functional sub-network constructed on the UK Bioabnk dataset using the SFCN algorithm has high brain age prediction accuracy and shows good generalization on independent datasets. Previous brain age-related studies have not yet constructed a brain age prediction model based on the functional sub-network. The MAE of the whole brain age prediction model is 3.5 years, and the coefficient of determination R is 1. 2 The MAE for the seven functional sub-networks was 0.67, and the MAEs for the seven functional sub-networks reached 4.4, 4.4, 4.7, 4, 4.2, 4.7, and 4.1 years, respectively. The brain age prediction model based on the ScaledDense U-Net algorithm achieved higher prediction accuracy than the SFCN model on the test set, with an MAE of 2.9 years and a Pearson correlation coefficient of 0.88 between predicted brain age and actual age.

[0071] In the healthy subjects of the test set, a significant negative correlation was found between cognitive scores such as fluid intelligence, digital memory, digit-symbol substitution test, and matrix reasoning and whole-brain PAD. This suggests that as PAD increases during normal aging, a person's perception, memory, computational speed, and reasoning abilities decline. Therefore, PAD can be used as a biomarker for assessing brain health during brain aging. This study further found that the correlation between PAD and cognitive scores exhibited network-specific patterns across different functional subnetworks in healthy subjects. In addition, a significant positive correlation was found between glial fibrillary acidic protein deposition and PAD in normal subjects and those with subjective cognitive decline in the Xuanwu Hospital dataset, and that glial fibrillary acidic protein accumulation is associated with brain neurodegeneration, further demonstrating that PAD can be used as a biomarker for brain neurodegeneration.

[0072] When exploring the spatial distribution of brain age at the voxel level, it was found that the prefrontal lobe of the brain was "younger" than other areas, and there was no significant between-group difference in the whole-brain voxel level PAD between normal people and people with subjective cognitive decline, indicating that compared with normal people, the brain structure degeneration of people with subjective cognitive decline may be very small, and the existing brain age prediction model is difficult to capture the characteristics of these changes.

[0073] The present invention also provides an application of a method for constructing a multi-scale brain age prediction model based on magnetic resonance images, which is applied in the risk prediction of Alzheimer's disease. PAD is used as a biomarker for evaluating brain health status during brain aging. There are significant differences in voxel-level PAD between normal people and people with mild cognitive impairment, normal people and people with Alzheimer's disease, and people with mild cognitive impairment and people with Alzheimer's disease, and the areas with significant differences are distributed in brain regions other than the middle frontal gyrus, the superior temporal gyrus, and the orbital gyrus.

[0074] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention are still within the scope of protection of the technical solution of the present invention.

Claims

1. A method for constructing a multi-scale brain age prediction model based on magnetic resonance imaging, characterized by: It includes the following steps: (1) Data collection: The T1-weighted MRI dataset of healthy subjects was used as the training set, and the mixed T1-weighted MRI dataset was used as the test set. The mixed T1-weighted MRI dataset contained T1-weighted MRI images of normal subjects, patients with subjective cognitive decline, mild cognitive impairment, and Alzheimer's disease. (2) Data preprocessing: The original images were preprocessed using Freesurfer software, including head motion correction, non-uniform field intensity normalization, linear registration, and scalping. The images were then resampled to a voxel space with a resolution of 2 mm and a size of 128 × 128 × 128 using the QSIPrep software package. The Yeo functional partitioning template was then nonlinearly registered to the individual space and uniformly resampled to a resolution of 2 mm to extract seven functional subnetworks: the visual network, the somatomotor network, the dorsal attention network, the ventral attention network, the limbic network, the frontoparietal network, and the default mode network. (3) Constructing a whole-brain and functional subnetwork brain age prediction model based on the simple fully convolutional neural network (SFCN) method: The whole-brain T1-weighted magnetic resonance images of each subject in the training set with a resolution of 2 mm and the seven functional subnetwork images obtained by segmentation based on the Yeo template were used as the input of the 3D neural network. The input image size was 128 × 128 × 128. A 3D deep convolutional neural network model was constructed based on the SFCN method. A series of 3D convolutional layers were used to extract high-dimensional features from the image. Finally, a fully connected layer was used for regression to output the predicted brain age. (4) Constructing a voxel-level brain age prediction model based on the ScaledDense U-Net method: using a whole-brain T1-weighted magnetic resonance image with a size of 91×109×91 and a resolution of 2 mm registered to the MNI standard space as the input of the ScaledDense U-Net method model, and also using the gender tensor as the input of the model. The output of the model is a predicted brain age map, which is a T1-weighted magnetic resonance image with the same size and resolution as the input image; (5) Bias correction of brain age deviation: The linear regression method is used to correct the bias of brain age deviation PAD. The bias correction method relies on the linear regression model of PAD and actual age. The formula for PAD bias correction is as follows: （1） in Represents the real age, is the slope, is the intercept, It is PAD, with linear regression model fitting and Afterwards, the corrected brain age is obtained by subtracting this bias from the predicted brain age, using the following formula: （2） in represents the corrected brain age, represents the predicted brain age. Formula (2) is applied to the test set to obtain the corrected brain age on the test set. The corrected brain age is then subtracted from the predicted brain age to obtain the corrected PAD.

2. The method for constructing a multi-scale brain age prediction model based on magnetic resonance imaging according to claim 1, characterized in that: In step (1), the UK Biobank dataset was used as a training set to construct a brain age prediction model, and the Xuanwu Hospital dataset, ADNI, and BABRI datasets were used as independent datasets to evaluate the generalization ability of the brain age prediction model and cross-validate the PAD group differences between different cognitive groups.

3. The method for constructing a multi-scale brain age prediction model based on magnetic resonance imaging according to claim 2, characterized in that: In step (2), the T1-weighted magnetic resonance image that has been registered to the standard space and resampled to a resolution of 2 mm is subjected to Z-Score normalization. The normalization formula is as follows: （3） in, This is the image before standardization. is the mean grayscale value of all voxels in image z, is the standard deviation of the grayscale values ​​of all voxels in image z, The definition is as follows: （4） in, is the mask corresponding to the brain, Indicates the Grayscale value of the voxel.

4. The method for constructing a multi-scale brain age prediction model based on magnetic resonance images according to claim 3, characterized in that: In the step (3), the first five modules are composed of a 3×3×3 convolution layer, a batch normalization layer, an ELU activation function layer and a maximum pooling layer. The sixth block is composed of a 1×1×1 convolution layer, a batch normalization layer, an ELU activation function layer and an average pooling layer. The pooling kernel size of the maximum pooling layer and the average pooling layer is 2×2×2. The first six blocks are used for feature extraction; the seventh block is composed of two fully connected layers and a Dropout layer. The first fully connected layer outputs a tensor of length 16, which is then passed through the ELU activation function and the Dropout layer. The Dropout layer retains or discards a certain proportion of neurons according to a given probability. The output value of the discarded neurons is set to 0 during the forward propagation process. The last fully connected layer outputs the final predicted brain age.

5. The method for constructing a multi-scale brain age prediction model based on magnetic resonance images according to claim 4, characterized in that: In step (3), the model initializes the network model parameters with the_uniform, which is a normal distribution initializer that starts with a center of 0 and a standard deviation of The images are sampled from a truncated normal distribution, where fan_in is the number of input neurons in the weight tensor. The dropout ratio of the Dropout layer is set to 0.2, which means that neurons are randomly dropped with a probability of 20% to avoid overfitting. The batch size of the network model is set to 32, and 32 image samples are loaded in each round of training. The optimizer is the Adam optimizer with an initial learning rate of 0.

01. The loss function is the L1 loss function. If the loss function does not decrease after 120 consecutive iterations of the model training, the learning rate is automatically multiplied by a factor of 0.1 to obtain a smaller learning rate, and the iterative training continues. The maximum number of iterations is set to 500. The T1-weighted magnetic resonance images of the UK Biobank dataset were divided into training, validation, and test sets according to a 6:2:2 ratio. To reduce the impact of the model's prediction accuracy on the training, validation, and test sets due to the uneven age distribution of the subjects in the dataset, the dataset was randomly divided according to a ratio to ensure that the age distribution of the subjects in the training, validation, and test sets was roughly the same. The whole-brain brain age prediction model and the brain age prediction model of the seven functional subnetworks were trained separately.

6. The method for constructing a multi-scale brain age prediction model based on magnetic resonance images according to claim 5, characterized in that: In the step (4), the model presents a U-shaped structure, with the left part being the encoder part and the right part being the decoder part. The feature maps of the same size in the left and right symmetrical parts of the U-shaped structure are connected by jump connections, and the bottom-level features are fused with the high-level features of the corresponding positions during the upsampling process, thereby retaining more high-resolution information. The model first performs a three-dimensional convolution operation on the input image with a step size of 1, a convolution kernel size of 7×7×7, and a number of channels of 8, and is activated by ELU to obtain a feature map of size 81×99×81 and a number of channels of 8. It then passes through 4 dense layer blocks and maximum pooling operations in succession. Each time the dense layer block is passed, the number of channels of the feature map increases to 3 times the original number, and each maximum pooling operation reduces the size of the feature map to half the original number, thereby achieving downsampling. In order to consider the influence of gender factors on brain age prediction, the gender tensor is input into a fully connected layer and two 1×1×1 convolution layers. The output tensor is concatenated with the feature map after 5 dense layer blocks, and the output feature map has a number of 2200 channels. The decoder part gradually restores the spatial resolution by performing deconvolution and upsampling operations on the feature map. After each upsampling, the decoder feature map is spliced ​​with the high-level features of the same size corresponding to the left encoder. The spliced ​​tensor passes through the upper part of the dense layer. This operation is repeated four times and then the output feature map is upsampled to the same size as the original input. The predicted brain age map is output. The kernel size of the deconvolution is 2×2×2, and the size of the output feature map is twice that of the input. Each dense connection block consists of two asymmetric convolution AC blocks and two batch normalization blocks. The AC block uses four three-dimensional convolutional layers. The features output by the SE module are only concatenated with the features before the first AC block. The number of channels of the input feature tensor is n, and the number of channels of the final output feature tensor is 3n. The number of output feature channels of the upsampling part of the dense connection block is half of the number of input feature channels. The number of channels of the input feature tensor is halved after passing through a convolutional layer. The output of the convolutional layer is concatenated with the feature tensor output by the SE network block, and after concatenation, it continues to pass through a convolutional layer to obtain the final output feature tensor.

7. The method for constructing a multi-scale brain age prediction model based on magnetic resonance imaging according to claim 1, characterized in that: Its application in Alzheimer's disease risk prediction, PAD as a biomarker for assessing brain health during brain aging, there are significant differences in voxel-level PAD between normal people and people with mild cognitive impairment, normal people and people with Alzheimer's disease, and people with mild cognitive impairment and people with Alzheimer's disease, and the areas with significant differences are distributed in brain regions other than the middle frontal gyrus, the superior temporal gyrus, and the orbital gyrus.

Images