Multi-scale brain age prediction model construction method based on magnetic resonance image and application

Abstract

According to the multi-scale brain age prediction model construction method based on the magnetic resonance image and the application, the constructed brain age prediction model is higher in generalization and robustness, higher prediction precision is kept, the whole brain-sub-network-voxel brain age can be predicted, the predicted brain age has better interpretability in the physiological sense, and the brain age prediction accuracy is improved. The difference of brain ages among different sub-networks and a specific mode of PAD and cognition association are explored, the specific sub-network for regulating cognition is found, the difference mode of aging of different brain regions is seen from the voxel level, and the prediction performance of the model is superior to that of a current mainstream neural network model. The method comprises the following steps: (1) data collection; (2) data preprocessing; (3) constructing a whole-brain and functional sub-network brain age prediction model based on a simple full convolutional neural network SFCN method; (4) constructing a voxel level brain age prediction model based on a ScaledDense U-Net method; and (5) carrying out offset correction on the brain age deviation.

Description

Technical Field

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

[0002] The initial brain age prediction models were constructed based on traditional machine learning algorithms, mainly using information such as gray matter and white matter extracted from brain images as features. Commonly used machine learning algorithm 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 the above traditional machine learning algorithms have achieved good results in brain age prediction research, providing effective methods for exploring scientific issues related to brain age. Among them, Franke et al. used the T1w images of 410 healthy subjects collected from different sites in the IXI dataset as the training set. After dimensionality reduction using the principal component analysis method, an RVM model was trained to obtain a brain age prediction model. Then, 245 healthy subjects were used as the test set to predict their brain ages, and the model was applied 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, the average difference between the brain age and the actual age (BrainAGE) in the AD population reached 10, and it was found that the training sample size was an important factor in the accuracy of the imaging model. Cole et al. extracted GM and WM from the T1w images of 1537 healthy subjects as input features, 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. Moreover, compared with normal subjects, the brains of TBI patients were "older". The differences between the predicted age and the actual age (PAD) in WM and GM were 5.97 and 4.66 respectively, indicating that the brain undergoes accelerated aging throughout the late stage of chronic injury. Huang et al. used Elastic Net to train a brain age prediction model with the gray matter characteristics of 974 healthy subjects (490 from the BABRI dataset and 484 from the ADNI dataset), and predicted the brain ages and PADs of 231 healthy subjects and 224 patients with aMCI. The results showed that the PADs of aMCI patients were higher than those of the normal control group. When considering the AD risk-related allele APOE 4, the group carrying this gene showed a higher PAD than the group not carrying this gene. It was also found that the PADs of amyloid-positive aMCI patients were higher than those of amyloid-negative patients.

[0003] In the past few years, machine learning algorithms have promoted the development of brain age prediction and achieved many results, demonstrating their unique advantages. However, when traditional machine learning algorithms are applied in brain age prediction research, it is necessary to manually extract features from brain images, and only after extracting meaningful features can the subsequent training process be carried out. In recent years, with the development of deep learning, various deep learning algorithm models have been applied to brain age prediction research. Compared with traditional machine learning algorithms, deep learning algorithms have complex models, can capture higher-level feature expressions from raw data, and can handle larger-scale data. Therefore, deep learning algorithms are favored by researchers and have promoted great progress in brain age prediction research. Jonsson et al. extracted Jacobian images, gray matter images, and white matter images from the brain T1w images of 1264 healthy people from Iceland, and used a 3D Convolutional Neural Network (CNN) to train the model. The final brain age was obtained by linearly combining the predicted four brain ages. In this study, the CNN model introduced a shortcut mechanism to alleviate the vanishing gradient phenomenon caused by deepening the network. At the same time, data augmentation techniques were used to prevent model overfitting. The trained brain age prediction model was tested on the UK Biobank dataset and transferred to a new dataset for learning, improving the model's generalization and prediction accuracy. The prediction results showed an MAE of 3.5 years. After performing a genome-wide association analysis on the predicted PAD, two variant sites were found. One was related to a reduction in sulcus width, and the other was related to a reduction in white matter surface area. To overcome the requirements of deep learning for large samples and computing resources, Peng et al. proposed a Simple Fully Convolutional Network (SFCN). Compared with other mainstream deep learning networks, this model has fewer parameters and is more suitable for large-sample datasets. Using the original T1w images as input features, a brain age model was trained in 14,503 subjects from the UK Biobank, and the predicted MAE reached 2.14 years, and the accuracy of gender classification was as high as 99.5%. Most brain age prediction studies focus on middle-aged and elderly populations. To study the brain age across all ages, Bashyam et al. used the DeepBrainNet algorithm model to train a brain age prediction model on a dataset consisting of 11,729 subjects from different datasets with an age range from 3 to 95 years. The results showed that compared with overfitted models, moderately fitted brain age prediction models could capture more significant differences in different disease groups. At the same time, it was also proved that transfer learning of the DeepBrainNet-based brain age prediction model could construct a more accurate brain disease classifier.Most of the current brain age prediction studies construct brain age prediction models based on the whole brain or regions of interest (ROIs). Although some models in brain age research have generated images of regions sensitive to brain age prediction and found some physiological and anatomical significance from the distribution range of voxels that contribute the 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, with T1w images as input features, trained a brain age prediction model on a dataset of 3,463 healthy individuals, and generated 3D brain age maps for individuals. It was found that the median of the mean absolute error of individual brain age was 9.5 years. After correcting for PAD at the voxel level, distinct local brain age patterns were found between the healthy control group 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 3D U-Net network to predict brain age at the voxel level of the brain on T1w brain images, obtained the brain age of the brain anatomical structure through a 3D segmentation mask, and calculated the deviation during 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 of the whole brain is output. The physiological significance and application value of the brain age estimation value provided by the brain age prediction model are poor, and the prediction accuracy is not high. Summary of the Invention

[0005] To overcome the defects of the prior art, 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, can predict the brain age of the whole brain - sub-network - voxel, makes the predicted brain age have better interpretability in terms of physiological significance, explores the differences in brain age between different sub-networks and the specific patterns of the association between its PAD and cognition, finds the specific sub-networks that regulate cognition, and from the voxel level, observes the differential patterns of aging in different brain regions. The prediction performance of the model is superior to that of the current mainstream neural network models.

[0006] The technical solution of the present invention is as follows: This method for constructing a multi-scale brain age prediction model based on magnetic resonance images includes the following steps: (1) Data collection: The T1-weighted magnetic resonance image dataset of healthy subjects is used as the training set, and the mixed T1-weighted magnetic resonance image dataset is used as the test set. The mixed T1-weighted magnetic resonance image dataset includes T1-weighted magnetic resonance images of normal people, subjective cognitive decline, mild cognitive impairment, and Alzheimer's disease patients; (2)Data preprocessing: The original images were preprocessed using Freesurfer software, including motion correction, non-uniform field intensity normalization, linear registration, scalp stripping, and then resampled to a voxel space with a resolution of 2 mm and a size of 128×128×128 using the QSIPrep software package. Then, the Yeo functional parcellation template was nonlinearly registered to the individual space and uniformly resampled to a 2 mm resolution for extracting seven functional subnetworks of the visual network, somatic motor network, dorsal attention network, ventral attention network, limbic network, fronto-parietal network, and 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 2 mm resolution whole-brain T1-weighted magnetic resonance images of each subject in the training set, as well as the seven functional subnetwork images segmented based on the Yeo template, were used as the inputs of the 3D neural network. The size of the input images was 128×128×128. A 3D deep convolutional neural network model was constructed based on the SFCN method. High-dimensional features in the images were extracted through a series of 3D convolutional layers, and finally, a fully connected layer was used for regression to output the predicted brain age; (4)Construct a voxel-level brain age prediction model based on the ScaledDense U-Net method: Use the whole-brain T1-weighted magnetic resonance images registered to the MNI standard space with a size of 91×109×91 and a resolution of 2 mm 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. The left part is the encoder part, and the right part is the decoder part. The feature maps of the same size in the symmetric parts of the U-shaped structure are connected by skip connections. During the upsampling process, the low-level features are fused with the high-level features at the corresponding positions to retain 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 passes through the ELU activation to obtain a feature map with a size of 81×99×81 and 8 channels. Then it successively passes through 4 dense layer blocks and max-pooling operations. Each time it passes through a dense layer block, the number of channels of the feature map increases to 3 times the original, and each max-pooling operation reduces the size of the feature map to half of the original to achieve downsampling. 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, and the output tensor is concatenated with the feature map after passing through 5 dense layer blocks to output a feature map with 2200 channels. The decoder part gradually restores the spatial resolution by performing deconvolution operations and upsampling operations on the feature map. After each upsampling, the decoder feature map is concatenated with the high-level features of the same size corresponding to the encoder in the left part. After such operations are repeated four times, the output feature map is upsampled once more to the same size as the original input to output the predicted brain age map, 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. (5)Bias correction of brain age deviation: Use the method of linear regression to perform bias correction on the brain age deviation PAD. The bias correction method depends on the linear regression model of PAD and the actual age. The formula for PAD bias correction is as follows: (1) where represents the true age, is the slope, is the intercept, is PAD. After obtaining and fitted by the linear regression model, subtract this bias from the predicted brain age to obtain the corrected brain age. The formula is as follows: (2) where represents the corrected brain age, represents the predicted brain age, and are the slope and intercept of the linear model fitted on the training set. Applying formula (2) to the test set gives the corrected brain age on the test set, and then subtracting the predicted brain age from the corrected brain age gives the corrected PAD.

[0007] The beneficial technical effects of the present invention are as follows: 1) Construct a brain age prediction model on a large sample dataset, making the model have better generalization. The current brain age prediction research uses a sample size that is not large enough for modeling. This study constructs a brain age prediction model on a sample of about 28,000 people in the UK Biobank dataset. The generalization and robustness of the model are stronger, enabling the model to maintain a high prediction accuracy when applied to other independent datasets.

[0008] 2) Construct a multi-scale brain age prediction model based on the whole brain, functional sub-networks, and voxels. Compared with the brain age model based on the whole brain, the multi-scale brain age model in this study can predict the brain age of "whole brain - sub-network - voxel", making the predicted brain age have better interpretability in terms of physiological significance.

[0009] 3) Construct 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 the association between their PAD and cognition, and find the specific sub-networks that regulate cognition.

[0010] 4) Construct a voxel-level brain age prediction model based on ScaledDense U-Net to view the differential patterns of aging in different brain regions at the voxel level. ScaledDense U-Net introduces deep supervision and ranking loss functions, and the prediction performance of the model is better than that of the current mainstream neural network models.

[0011] Also provided is the application of the method for constructing a multi-scale brain age prediction model based on magnetic resonance images. Its application in Alzheimer's disease risk prediction shows that PAD, as a biomarker for evaluating the brain health status during the brain aging process, there are significant differences in voxel-level PAD between normal people and people with mild cognitive impairment, normal people and Alzheimer's patients, and people with mild cognitive impairment and Alzheimer's patients. Moreover, the regions with significant differences are distributed in brain regions other than the middle frontal gyrus of the brain, the superior temporal gyrus of the brain, and the orbital gyrus of the brain. Description of the Drawings

[0012] Figure 1 is a flowchart of the method for constructing a multi-scale brain age prediction model based on magnetic resonance images according to the present invention. Detailed Embodiments

[0013] Such as Figure 1As shown, this method for constructing a multi-scale brain age prediction model based on magnetic resonance images includes the following steps: (1) Data collection: The T1-weighted magnetic resonance image dataset of healthy subjects is used as the training set, and the mixed T1-weighted magnetic resonance image dataset is used as the test set. The mixed T1-weighted magnetic resonance image dataset contains T1-weighted magnetic resonance images of normal people, subjective cognitive decline, mild cognitive impairment, and Alzheimer's disease patients; (2) Data preprocessing: Use Freesurfer software to preprocess the original images, including head motion correction, non-uniform field intensity normalization, linear registration, scalp stripping, and then resample to a voxel space with a resolution of 2mm and a size of 128×128×128 through the QSIPrep software package. Then, non-linearly register the Yeo functional partition template to the individual space and uniformly resample to a 2mm resolution for extracting seven functional sub-networks of the visual network, somatic motor network, dorsal attention network, ventral attention network, limbic network, fronto-parietal network, and default mode network; (3) Construct a whole-brain and functional sub-network brain age prediction model based on the simple fully convolutional neural network (SFCN) method: Use the whole-brain T1-weighted magnetic resonance images with a resolution of 2mm for each subject in the training set and the seven functional sub-network images segmented based on the Yeo template as the input of the 3D neural network. The input image size is 128×128×128. Construct a 3D deep convolutional neural network model based on the SFCN method, extract high-dimensional features in the images through a series of 3D convolutional layers, and finally use a fully connected layer for regression to output the predicted brain age; (4)Construct a voxel-level brain age prediction model based on the ScaledDense U-Net method: Use the whole-brain T1-weighted magnetic resonance images registered to the MNI standard space with a size of 91×109×91 and a resolution of 2mm 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. The left part is the encoder part, and the right part is the decoder part. The feature maps of the same size in the symmetric parts on the left and right of the U-shaped structure are connected by skip connections. During the upsampling process, the underlying features are fused with the high-level features at the corresponding positions to retain 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 passes through the ELU activation to obtain a feature map with a size of 81×99×81 and 8 channels. Then it successively passes through 4 dense layer blocks and max pooling operations. Each time it passes through a dense layer block, the number of channels of the feature map increases to 3 times the original, and each max pooling operation reduces the size of the feature map to half of the original to achieve downsampling. 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, and the output tensor is concatenated with the feature map after passing through 5 dense layer blocks to output a feature map with 2200 channels. The decoder part gradually restores the spatial resolution through deconvolution operations and upsampling operations on the feature map. After each upsampling, the decoder feature map is concatenated with the high-level features of the same size corresponding to the encoder on the left side. After such operations are repeated four times, the output feature map is upsampled once more to the same size as the original input to output the predicted brain age map, 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. (5)Bias correction of brain age deviation: Use the method of linear regression to perform bias correction on the brain age deviation PAD. The bias correction method depends on the linear regression model of PAD and the actual age. The formula for PAD bias correction is as follows: (1) where represents the real age, is the slope, is the intercept, is PAD. After obtaining the and fitted by the linear regression model, subtract this bias from the predicted brain age to obtain the corrected brain age. The formula is as follows: (2) where represents the corrected brain age, represents the predicted brain age, and are the slope and intercept of the linear model fitted on the training set. Applying formula (2) to the test set gives the corrected brain age on the test set, and then subtracting the predicted brain age from the corrected brain age gives the corrected PAD.

[0014] The beneficial technical effects of the present invention are as follows: 1) Construct a brain age prediction model on a large-sample dataset, making the model have better generalization. The current brain age prediction research uses a sample size that is not large enough for modeling. In this study, a brain age prediction model is constructed on a sample of about 28,000 people in the UK Biobank dataset. The generalization and robustness of the model are stronger, enabling the model to maintain a high prediction accuracy when applied to other independent datasets.

[0015] 2) Construct a multi-scale brain age prediction model based on the whole brain, functional subnetworks, and voxels. Compared with the brain age model based on the whole brain, the multi-scale brain age model in this study can predict the brain age of "whole brain - subnetworks - voxels", making the predicted brain age have better interpretability in terms of physiological significance.

[0016] 3) Construct brain age prediction models based on seven subnetworks respectively, explore the differences in brain age between different subnetworks and the specific patterns of the association between their PAD and cognition, and find the specific subnetworks that regulate cognition.

[0017] 4) Construct a voxel-level brain age prediction model based on ScaledDense U-Net to view the differential patterns of aging in different brain regions at the voxel level. ScaledDense U-Net introduces deep supervision and ranking loss functions, and the prediction performance of the model is better than that of the current mainstream neural network models.

[0018] Preferably, in step (1), using the UK Biobank dataset as the training set to construct a brain age prediction model, and using the Xuanwu Hospital dataset, ADNI, and BABRI datasets as independent datasets to evaluate the generalization ability of the brain age prediction model and cross-validate the between-group differences in PAD among different cognitive groups.

[0019] Preferably, in step (2), perform Z-Score normalization on the T1-weighted magnetic resonance images that have been registered to the standard space and resampled to a resolution of 2 mm. The normalization formula is as follows: (3) where is the image before normalization, is the image the mean of the gray values of all voxels, is the standard deviation of the gray values of all voxels in the image y. The standard deviation is defined as follows: (4) where is the value of the corresponding voxel in the brain mask, represents the gray value of the individual voxel.

[0020] Preferably, in the step (3), the first five modules are all composed of a 3×3×3 convolutional layer, a batch normalization layer, an ELU activation function layer, and a max pooling layer. The sixth block is composed of a 1×1×1 convolutional layer, a batch normalization layer, an ELU activation function layer, and an average pooling layer. The pooling kernel sizes of the max pooling layer and the average pooling layer are both 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 with a length of 16. Then, it passes through an ELU activation function and a Dropout layer. The Dropout layer retains or discards a certain proportion of neurons according to a given probability. The discarded neurons have their output values set to 0 during the forward propagation process. The last fully connected layer outputs the final predicted brain age.

[0021] Preferably, in the step (3), the model initializes the network model parameters with the_uniform. The_uniform is a normal distribution initializer that samples from a truncated normal distribution centered at 0 with a standard deviation of 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, that is, neurons are randomly discarded 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 and a weight decay coefficient of 1×10 -6 ; The loss function is the L1 loss function. When the loss function of the model does not decrease for 120 consecutive iterations during training, the learning rate is automatically multiplied by a factor of 0.1 to obtain a smaller learning rate, and training continues iteratively. The maximum number of iterations is set to 500 times; The T1-weighted magnetic resonance images of the UK Biobank dataset are divided into a training set, a validation set, and a test set according to a ratio of 6:2:2. To reduce the bias in the prediction accuracy of the model on the training set, validation set, and test set caused by the uneven age distribution of the subjects in the dataset, when dividing the dataset, it is randomly divided according to a ratio and it is ensured that the age distributions of the subjects in the training set, validation set, and test set are approximately the same. The whole-brain brain age prediction model and the brain age prediction models of seven functional sub-networks are trained respectively.

[0022] Preferably, in the step (4), each dense connection block is composed of two asymmetric convolution AC blocks, two batch normalization layers, an ELU activation function layer, and a squeeze-and-excitation SE module. The AC block uses four three-dimensional convolution 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 finally output feature tensor is 3n. The number of output feature channels in 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 convolution layer. The output of the convolution layer is concatenated with the feature tensor output by the SE network block, and the final output feature tensor is obtained after passing through another convolution layer.

[0023] Preferably, in the step (4), for N samples, the Spearman rank correlation coefficient SRCC is defined as the Spearman rank correlation coefficient between the ranked values of two variables: (5) where Rank(·) is the ranking operation, is the covariance of two sets of ranked values, is the standard deviation of two sets of ranked values. If and have no identical values, the above definition is equivalent to: (6) If there are identical values in two sets of variables, their ranked values take the mean rank of these identical values in the increasing sequence; The ranking loss function is defined as follows: (7) Use the SoDeep model to replace the Rank(·) operation. SoDeep trains a differentiable proxy network R based on synthetic data, where 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: (8) where R(·) is an approximate representation of the variable ranking. SoDeep trains the network R until the model converges by minimizing ; Calculate the mean squared error MSE loss function for the feature maps output by the first, second, and third layers in the decoder part of the ScaledDense U-Net network respectively, to obtain the loss functions and as well as , and the MSE loss function between the finally output brain age map of the model and the target value is , and the ranking loss function is , the final loss function is the linear weighted sum of these four loss functions, and the formula is as follows: (9) where , , , and are the weight coefficients for each loss function.

[0024] Preferably, in the step (4), the finally output brain age map in the network is subjected to a voxel-level smoothing process once, the smoothing kernel size is 3×3×3, the loss function during 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 during the training process. This 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 , and the learning rate update strategy is that when the loss function on the validation set does not decrease for 10 epochs, the learning rate is multiplied by 0.5 to obtain a new learning rate, and the model is continued to be trained until the loss function on the validation set still cannot decrease after 30 consecutive epochs, then the training of the model is stopped.

[0025] Preferably, this method further includes step (6), model evaluation, and the evaluation indicators include the Pearson correlation coefficient, the coefficient of determination, and the mean absolute error MAE between the predicted brain age and the actual age; The definition of the Pearson correlation coefficient is as follows: (10) The coefficient of determination reflects the proportion of the variation of the dependent variable explained by the independent variable through the regression relationship, and its definition is as follows: (11) The mean absolute error MAE is as follows: (12) where is the brain age map predicted by this model, is the mask of the brain, is the average value of the brain age at the voxel level of the whole brain, and this average brain age value is used as the predicted brain age value in all subsequent statistical analyses based on the individual whole brain.

[0026] From the results of the present invention, it is found that the brain age prediction model of the whole brain and functional sub-networks constructed using the SFCN algorithm on the UK BioBank dataset has relatively high brain age prediction accuracy and shows good generalization on independent datasets, while previous brain age-related studies have not constructed brain age prediction models on functional sub-networks. The MAE of the whole brain brain age prediction model is 3.5 years, and the coefficient of determination R 2 is 0.67. The MAEs on the seven functional sub-networks reach 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 achieves higher prediction accuracy than the SFCN model on the test set, with an MAE of 2.9 years and a Pearson correlation coefficient between the predicted brain age and the actual age of 0.88.

[0027] In the healthy subjects of the test set, it is found that there are significant negative correlations between cognitive scores such as fluid intelligence, digit memory, digit-symbol substitution test, and matrix reasoning and the whole brain PAD. This indicates that during the normal aging process, as the PAD increases, people's perception, memory, operation speed, and reasoning ability will decline. Therefore, PAD can be used as a biomarker for evaluating the brain health status during the brain aging process. This study further finds that on different functional sub-networks of healthy subjects, the correlation between PAD and cognitive scores shows a network-specific pattern. In addition, in the normal and subjective cognitive decline populations of the Xuanwu Hospital dataset, it is found that there is a significant positive correlation between the deposition of glial fibrillary acidic protein and PAD, and the accumulation of glial fibrillary acidic protein is related to brain neurodegeneration, further proving that PAD can be used as a biomarker for brain neurodegeneration.

[0028] When exploring the spatial distribution of brain age at the voxel level, it is found that the prefrontal part of the brain is relatively "younger" compared to other regions, and there is no significant between-group difference in the whole brain voxel-level PAD between normal and subjective cognitive decline populations, indicating that compared with the normal population, the brain structure degeneration of the subjective cognitive decline population may be very small, and the existing brain age prediction models are difficult to capture the characteristics of these changes.

[0029] The present invention also provides an application of the method for constructing a multi-scale brain age prediction model based on magnetic resonance images, which is applied to the prediction of Alzheimer's disease risk. As a biomarker for evaluating the brain health status during the brain aging process, there are significant differences in the voxel-level PAD between normal and mild cognitive impairment populations, normal and Alzheimer's disease populations, and mild cognitive impairment and Alzheimer's disease populations, and the regions with significant differences are distributed in brain regions other than the middle frontal gyrus, superior temporal gyrus, and orbital gyrus of the brain.

[0030] The above are only the preferred embodiments of the present invention, and do not impose any form of limitation on the present invention. Any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention still fall within the protection scope 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 images, characterized in that: It includes the following steps: (1) Data collection: The T1-weighted magnetic resonance image dataset of healthy subjects is used as the training set, and the mixed T1-weighted magnetic resonance image dataset is used as the test set. The mixed T1-weighted magnetic resonance image dataset contains T1-weighted magnetic resonance images of normal people, subjective cognitive decline, mild cognitive impairment, and Alzheimer's disease patients; (2) Data preprocessing: The original images are preprocessed using Freesurfer software, including head motion correction, non-uniform field intensity normalization, linear registration, and scalp stripping. Then, through the QSIPrep software package, the images are resampled to a voxel space with a resolution of 2 mm and a size of 128×128×128. Then, the Yeo functional parcellation template is non-linearly registered to the individual space and uniformly resampled to a 2 mm resolution for extracting seven functional sub-networks of the visual network, somatomotor network, dorsal attention network, ventral attention network, limbic network, fronto-parietal network, and default mode network; (3) Construct a whole-brain and functional sub-network brain age prediction model based on the simple fully convolutional neural network (SFCN) method: The whole-brain T1-weighted magnetic resonance images with a resolution of 2 mm of each subject in the training set and the seven functional sub-network images segmented based on the Yeo template are used as the inputs of the 3D neural network. The input image size is 128×128×128. A 3D deep convolutional neural network model is constructed based on the SFCN method. High-dimensional features in the images are extracted through a series of 3D convolutional layers, and finally, a fully connected layer is used for regression to output the predicted brain age; (4) Construct a voxel-level brain age prediction model based on the ScaledDense U-Net method: The 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 is used as the input of the ScaledDense U-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. The left part is the encoder part, and the right part is the decoder part. Feature maps of the same size in the symmetric parts on the left and right of the U-shaped structure are connected by skip connections. During the upsampling process, the underlying features are fused with the high-level features at the corresponding positions to retain 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 passes through the ELU activation to obtain a feature map with a size of 81×99×81 and 8 channels. Then, it successively passes through 4 dense layer blocks and max-pooling operations. Each time it passes through a dense layer block, the number of channels of the feature map increases to 3 times the original, and each time the max-pooling operation makes the size of the feature map become half of the original to achieve downsampling; 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 convolutional layers, and the output tensor is concatenated with the feature map after passing through 5 dense layer blocks to output a feature map with 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 concatenated with the high-level feature of the same size corresponding to the left encoder part. After the concatenated tensor passes through the upper part of the dense layer, such operations are repeated four times, and then the output feature map is upsampled once to the same size as the original input to output the predicted brain age map. Among them, the kernel size of the deconvolution is 2×2×2, and the size of the output feature map obtained is twice that of the input. (5) Bias correction of brain age deviation: Use the method of linear regression to perform bias correction on the brain age deviation PAD. The bias correction method depends on the linear regression model of PAD and the actual age. The formula for PAD bias correction is as follows: （1） Wherein represents the real age, is the slope, is the intercept, is the PAD. After the linear regression model is fitted with and , the predicted brain age is subtracted by this bias to obtain the corrected brain age. The formula is as follows: （2） where represents the corrected brain age, represents the predicted brain age, and are the slope and intercept of the linear model fitted on the training set. Applying formula (2) to the test set gives the corrected brain age on the test set, and then subtracting the predicted brain age from the corrected brain age gives the corrected PAD.

2. The method for constructing a multi-scale brain age prediction model based on magnetic resonance images according to claim 1, wherein: In step (1), the UK Biobank dataset is used as the training set to construct a brain age prediction model. 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 between-group differences of PAD among different cognitive groups.

3. The method for constructing a multi-scale brain age prediction model based on magnetic resonance images according to claim 2, wherein: In step (2), the T1-weighted magnetic resonance images that have been registered to the standard space and resampled to 2mm resolution are further processed by Z-Score normalization. The formula for normalization is as follows: （3） Among them, is the image before standardization, is the image the mean of all voxel gray values, is the standard deviation of all voxel gray values of image y, and the standard deviation is defined as follows: （4） Among them, is the value of the corresponding voxel in the brain mask, representing the gray 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, wherein: In step (3), the first five modules are all composed of a 3×3×3 convolutional layer, a batch normalization layer, an ELU activation function layer, and a max pooling layer. The sixth block is composed of a 1×1×1 convolutional layer, a batch normalization layer, an ELU activation function layer, and an average pooling layer. The pooling kernel sizes of the max pooling layer and the average pooling layer are both 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 with a length of 16. Then, it passes 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 discarded neurons have their output values 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, wherein: In the step (3), the network model parameters are initialized with the_uniform, where the_uniform is a normal distribution initializer that samples from a truncated normal distribution centered at 0 with a standard deviation of , 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, 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 for each round of training; the optimizer is the Adam optimizer, the initial learning rate is 0.01, and the weight decay coefficient is 1×10 -6 ; The loss function is the L1 loss function. When the loss function does not decrease continuously for 120 iterations during the training process of the model, the learning rate is automatically multiplied by a factor of 0.1 to obtain a smaller learning rate, and the training continues iteratively. The maximum number of iterations is set to 500 times. The T1-weighted magnetic resonance images of the UK Biobank dataset are divided into a training set, a validation set, and a test set according to a ratio of 6:2:

2. In order to reduce the bias impact on the prediction accuracy of the model on the training set, validation set, and test set caused by the uneven age distribution of the subjects in the dataset, when dividing the dataset, it is randomly divided according to the ratio and it is ensured that the age distribution of the subjects in the training set, validation set, and test set is roughly the same. The brain age prediction models of the whole brain and seven functional subnetworks are trained respectively.

6. The method for constructing a multi-scale brain age prediction model based on magnetic resonance images according to claim 5, wherein: In the step (4), each dense connection block consists of two asymmetric convolution AC blocks, two batch normalization layers, an ELU activation function layer, and a squeeze-and-excitation SE module. The AC block uses four 3D convolution 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 finally output feature tensor is 3n. The number of output feature channels in 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 convolution layer, and the output of the convolution layer is concatenated with the feature tensor output by the SE network block. After concatenation, it continues to pass through a convolution layer to obtain the final output feature tensor.

7. The method for constructing a multi-scale brain age prediction model based on magnetic resonance images according to claim 6, wherein: In the step (4), for N samples, the Spearman rank correlation coefficient SRCC is defined as the Spearman rank correlation coefficient between the ranked values of two variables: （5） Among them Rank(·) is a sorting operation is the covariance of two sets of sorting values is the standard deviation of two sets of sorting values. If and have no identical values, the above definitions are equivalent to: （6） If there are the same values in two sets of variables, their ranked values take the mean of the ranks of these same values in the increasing sequence; The ranking loss function is defined as follows: （7） Use the SoDeep model to replace the Rank(·) operation. SoDeep trains a differentiable proxy network R based on synthetic 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: （8） where R(·) is an approximate representation of the variable sorting, and SoDeep trains the network R until the model converges by minimizing ; Calculate the mean squared error (MSE) loss function for the feature maps output by the first, second, and third layers of the decoder part in the ScaledDense U-Net network respectively to obtain the loss function and as well as . The MSE loss function between the brain age map finally output by the model and the target value is , the ranking loss function is . The final loss function is the linear weighted sum of these four loss functions, and the formula is as follows: （9） Among them , , , and are the weight coefficients before each loss function.

8. The method for constructing a multi-scale brain age prediction model based on magnetic resonance images according to claim 7, wherein: In the step (4), a voxel-level smoothing process is performed on the finally output brain age map in the network. The size of the smoothing kernel is 3×3×3. The loss function during 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 during the training process. This 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 , and the learning rate update strategy is that when the loss function on the validation set does not decrease for 10 epochs, the learning rate is multiplied by 0.5 to obtain a new learning rate, and the model is continued to be trained until the loss function on the validation set still cannot be decreased after 30 consecutive epochs, then the training of the model is stopped.

9. The method for constructing a multi-scale brain age prediction model based on magnetic resonance images according to claim 7, wherein: This method further includes step (6), model evaluation. The evaluation metrics include the Pearson correlation coefficient, the coefficient of determination, and the mean absolute error MAE between the predicted brain age and the actual age; The definition of the Pearson correlation coefficient is as follows: （10） The coefficient of determination reflects the proportion of the variation of the dependent variable explained by the independent variable through the regression relationship, and its definition is as follows: （11） The mean absolute error MAE is as follows: （12） Among them is the brain age map predicted by the model, is the mask of the brain, is the average value of the brain age at the voxel level of the whole brain. This average brain age value is used as the predicted brain age value in all subsequent statistical analyses based on the individual's whole brain.

10. Use of the method for constructing a multi-scale brain age prediction model based on magnetic resonance images according to claim 1, characterized in that: It is applied in Alzheimer's disease risk prediction. As a biomarker for evaluating the brain health status during the brain aging process, there are significant differences in voxel-level PAD between normal people and people with mild cognitive impairment, normal people and Alzheimer's patients, and people with mild cognitive impairment and Alzheimer's patients. Moreover, the regions with significant differences are distributed in the brain regions of the non-middle frontal gyrus, non-superior temporal gyrus, and non-orbital gyrus of the brain.

Images