Brain age estimation method based on adversarial learning

Abstract

The application discloses a brain age estimation method based on adversarial learning, extracts corresponding image features specific to age based on whole brain structural T1 magnetic resonance imaging, and realizes prediction of brain age, and comprises the following steps: step one: data collection and pretreatment, more than 2000 cases of data from 5 sites are collected, the age span is 5-94, the collected data is pretreated, all brain image data is registered to MNI standard space, and the size is standardized; step two: based on the adversarial learning deep neural network, the network is trained and optimized, and prediction of brain age is realized; step three: performance evaluation of brain age prediction task, optimization of prediction results, model feature extraction capability and generalization capability evaluation. Compared with the prior art, the application reduces the phenomenon that the model training result is unstable and the prediction effect is not ideal due to uneven age distribution of samples.

Description

Technical Field

[0001] The present invention relates to the technical field of medical intelligent image processing, and in particular to a brain age estimation method based on adversarial learning. Background Art

[0002] The brain, the second most important organ after the heart, has become a major threat to human health and well-being as living standards and hygiene improve. Therefore, measuring brain health and studying the developmental mechanisms of brain disease have become hot topics. MRI, the gold standard for diagnosing some brain diseases, is widely used in brain disease research. Numerous studies have shown that patients with various brain diseases exhibit varying degrees of structural and functional changes in the brain. For example, patients with Alzheimer's disease at different stages show significant reductions in gray matter volume, particularly in the hippocampus, while the mean diffusivity in white matter increases. Parkinson's disease patients exhibit strong T2 hyperintensities in the periventricular region, globus pallidus, and substantia nigra. Schizophrenia patients also exhibit significant reductions in gray matter in the frontal and prefrontal cortex regions, impairing functional memory, such as working and declarative memory. Although numerous studies have proposed different pathogenic mechanisms for various brain diseases, the complexity of brain function and structure makes it difficult to develop a reliable indicator to measure brain health.

[0003] Brain age has been proposed as a new marker to measure brain development. The difference between predicted brain age and actual age is used as an age marker to estimate brain aging and health. For example, studies have found that the brain age of patients with schizophrenia and other mental illnesses is younger than that of healthy controls. In patients with certain geriatric diseases, the observed brain age is significantly higher than that of age-matched controls. Therefore, accurately estimating brain age is crucial for studying brain development and aging, as well as related brain diseases. Currently, there are many methods for predicting brain age, including those based on structural imaging (T1 imaging, diffusion tensor imaging (DTI)) and functional magnetic resonance imaging (FMRI). However, due to the large individual variability and high noise levels of functional imaging, FMRI is far less reliable than structural imaging for predicting brain age. Therefore, the most commonly used method for predicting brain age is structural MRI. Methods for applying regression models to structural MRI brain imaging data fall into two main categories: common machine learning methods and various deep learning methods. Common machine learning models for predicting brain age include linear regression, support vector regression, and Gaussian process regression. However, these machine learning methods are significantly affected by the number of features. When the input features are large, dimensionality reduction methods are required to achieve accurate brain age prediction. For example, when using whole-brain gray matter voxels as input for brain age prediction, dimensionality reduction is required before inputting into a machine learning model. Deep learning algorithms, on the other hand, can ignore the input dimensionality issue and directly feed whole-brain MRI images into a deep neural network. Finally, a fully connected regression activation layer is applied to predict brain age. However, deep learning currently suffers from poor generalization, making it difficult to directly transfer predictions from one dataset to new data. Furthermore, because brain age prediction is based on MRI images, site effects (variations caused by different scanning devices and scanning parameters) exacerbate the transferability of deep neural network models. At the same time, brain development and the volume of different brain structures are different in men and women. Therefore, it is not reasonable to directly construct a comprehensive brain age prediction model and apply it to the brain age prediction of men and women. However, most of them still establish a prediction model to predict the brain age of the population. Simply adding gender as a feature in the bottleneck layer of the model cannot effectively predict the brain age of cross-gender people. Summary of the Invention

[0004] The purpose of the present invention is to provide a brain age estimation method based on adversarial learning to solve the above technical problems:

[0005] The purpose of the present invention can be achieved through the following technical solutions:

[0006] The brain age estimation method based on adversarial learning extracts age-specific imaging features based on whole-brain T1 magnetic resonance imaging to predict brain age. The method includes the following steps:

[0007] Step 1: Data collection and preprocessing: Collect data from more than 2,000 subjects across five sites, ranging in age from 5 to 94 years old. Preprocess the collected data and align all brain imaging data to the MNI standard space to standardize the size.

[0008] Step 2: Based on the adversarial learning deep neural network, the network is trained and optimized to predict brain age;

[0009] Step 3: Performance evaluation of the brain age prediction task, optimization of prediction results, model feature extraction capability, and generalization ability evaluation.

[0010] As a further solution of the present invention: the specific data collection and preprocessing method in step 1 includes:

[0011] We collected 2011 samples from five datasets: South lifespan, CBIC, AIBL, IXI, and OASIS;

[0012] All data were preprocessed using a unified pipeline. This included grayscale correction of the MRI structural brain images, skull stripping, and registration of the brain images to the MNI standard space using a nonlinear registration method.

[0013] Establish a data set and divide the preprocessed brain imaging data into training set, validation set, and test set in a ratio of 8:1:1.

[0014] As a further solution of the present invention: in step 2, the adversarial learning deep neural network includes a main task model and an adversarial learning task model, wherein the main task model includes two parts, namely an encoder part and a brain age prediction part. The method for building the main task model specifically includes the following steps:

[0015] The first part is the encoder, which consists of 6 convolution modules. Each convolution module contains a convolution layer, a normalization layer, a pooling layer, and an activation layer. The number of channels corresponding to the 6 convolution modules is 32, 64, 128, 256, 256, and 128 respectively.

[0016] The second part is the brain age prediction part. The brain age prediction includes two parts: sample optimization and regression calculation. The brain age is predicted based on the last linear activation layer, and the result is estimated based on the L1 loss function.

[0017] As a further solution of the present invention: in step 2, the specific steps of sample optimization in the brain age prediction part are as follows:

[0018] The corresponding age is mapped to the range of 5-94 with a normal distribution of expected value of 0 and mean value of 1. The mapping formula is as follows:

[0019]

[0020] Where y is the corresponding age label, μ and σ correspond to the expectation and variance of the normal distribution. For each hashed label, KL divergence is used to constrain the output of the result. The total loss function of this part of the task is:

[0021]

[0022] where y i It represents the true brain age. represents the estimated brain age, p yi (x j ) represents y i The hashed tag value.

[0023] As a further solution of the present invention: in the step 2, the method for establishing the adversarial learning task model includes the following steps:

[0024] To achieve the classification task of multiple mixed factors, that is, to use cross entropy to constrain the classification mixed factor task, the main loss function form is as follows:

[0025]

[0026] Where i is the number of mixing factor modules, C i Represents the output category of factor classification, d c represents the output of the classifier;

[0027] Based on the common loss function task, the loss function is minimized, that is, the encoder is allowed to extract the features related to the task to the maximum extent, and the encoder is allowed to extract the features related to the mixing factor to the minimum extent. A gradient reversal layer is added between the encoder and the mixing factor module to minimize the encoder's extraction of the mixing factor features. Therefore, the total loss function is:

[0028] Loss=Loss age +λLossc_f

[0029] Here, λ corresponds to the weight of the loss function of the mixing factor, and here we take 0.8.

[0030] As a further solution of the present invention, the specific training steps in the second step are as follows:

[0031] Input: training set D = {(x(n), y(n))}, validation set V, learning rate α, hyperparameter ε, regularization coefficient β of the loss function, encoder E i , mixing factor module M i , brain age prediction module BA i , adversarial task parameter con_st = 0.5;

[0032] Randomly initialize the network weights w i , b i ε = 0 (w i ∈w, b i ∈B)

[0033] Train and optimize through the neural network model to obtain (w, b);

[0034] The specific training process is as follows:

[0035] Set the total number of loops K;

[0036] Randomly reorder the samples in the training set D, divide the samples in D into N batches, and the number of samples in each batch is 8;

[0037] For the samples in the L-th batch, update the parameter ε = (i + L * i) / (N * k), L < N, and i is the current number of loops;

[0038] Select a sample (x i , y i ) from the training set D, calculate the output of the image features passing through E i , M i , BA i , and use the loss function to constrain the backpropagation calculation of the error δ of the gradient passed back by the mixing factor d , the gradient error δ of brain age prediction BA , to optimize the parameters of the encoder's gradient. The error of the reverse gradient is: <*

[0039]

[0040] Reverse-calculate the derivative of the parameters of each layer of the encoder according to the loss function; <000*0147>Update the parameters;

[0042] Perform the i+1th loop until the error rate of the neural network model on the validation set V no longer decreases, and then output (W, B), where W represents the weights of all neurons in the neural network and B represents the bias.

[0043] As a further solution of the present invention: In step 3, the specific method for evaluating the performance of the brain age prediction task is as follows:

[0044] Calculate the mean absolute error: Among them, y is the true label, Represents the predicted label. The lower the MAE, the higher the predicted age accuracy.

[0045] Calculate the correlation coefficient r:

[0046]

[0047] in, represents the mean of the predicted values, It represents the mean of the true values ​​and is used to evaluate the similarity between the distribution of the true value and the predicted value. The closer r is to 1, the better the prediction result. If it is greater than or equal to 0.8, it means that there is a greater similarity between the predicted brain age and the true age, and the prediction effect is good.

[0048] Calculate R square error:

[0049]

[0050] in, It represents the average value of the output, which is used to evaluate the effect of the model fitting the data. The closer it is to 1, the better the model fitting effect is. 2 The lower it is, the worse the model fitting effect is.

[0051] As a further solution of the present invention: the step 3 also includes brain age adaptive correction, by setting the output brain age prediction threshold Th max =7, to correct samples whose predicted brain age exceeds the threshold. The threshold for brain age prediction is the difference between the actual brain age and the predicted brain age. The specific steps of the brain age correction method are as follows:

[0052] Establish a linear fitting model:

[0053] p a =α×Ω+β

[0054] Where Ω is the real age, p a Represents the predicted brain age, α and β are obtained by fitting the predicted brain age and the actual age data on the test set;

[0055] Correction by interpolation:

[0056] C a =p a +[Ω-(α×Ω+β)]

[0057] Among them C a It is the corrected brain age.

[0058] This invention focuses on optimizing the poor generalization of brain age prediction models and reducing biases due to gender differences and differences in scanning machine parameters. This facilitates the establishment of a unified end-to-end brain age prediction model to predict brain age in people from different regions. The main advantages of this invention are as follows:

[0059] 1. Based on the basic brain age regression prediction task constraints, we integrated a label hashing loss function to reduce the instability of model training results and unsatisfactory prediction results caused by the uneven age distribution of samples.

[0060] 2. The training idea of ​​adversarial learning is introduced to shield the features related to the mixing factors during the training process, ensuring the reliability of the extracted features and greatly improving the generalization results on other independent data sets.

[0061] 3. During the training process, the dynamic changes of the adversarial parameters indirectly add noise to the features of the feature encoder, greatly improving the stability of the feature extractor and improving the compatibility of brain MRI images with different preprocessing processes. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] The present invention will be further described below with reference to the accompanying drawings.

[0063] Figure 1 1 is a flow chart of the brain age estimation method based on adversarial learning according to the present invention;

[0064] Figure 2 These are the validation results of the adversarial learning-based brain age estimation method of the present invention on an external multi-site dataset (Autism Brain Imaging Data Exchange, a dataset from six sites). A and C are the distribution results of the model without the mixing factor module and the adversarial learning strategy, after the encoder output is mapped to a high-dimensional space (TSNE); B and D are the outputs of the model with the mixing factor module and adversarial learning applied.

[0065] Figure 3This is a performance comparison of the brain age estimation method based on adversarial learning in the present invention with various traditional machine learning models and deep learning models on a test set, where GM is gray matter, WM is white matter, and ALL brain is the registered whole-brain image as input;

[0066] Figure 4 Verification results of deep neural networks and deep neural networks using adversarial learning and mixing factors on external independent data sets (ABIDE, UK biobank). DETAILED DESCRIPTION

[0067] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0068] See also Figure 1 As shown, the present invention is a brain age estimation method based on adversarial learning. Based on the whole-brain T1 magnetic resonance imaging of the brain structure, the corresponding age-specific image features are extracted to achieve brain age prediction, which includes the following steps:

[0069] Step 1: Data collection and preprocessing: Collect data from more than 2,000 subjects across five sites, ranging in age from 5 to 94 years old. Preprocess the collected data and align all brain imaging data to the MNI standard space to standardize the size.

[0070] Step 2: Based on the adversarial learning deep neural network, the network is trained and optimized to predict brain age;

[0071] Step 3: Performance evaluation of the brain age prediction task, optimization of prediction results, model feature extraction capability, and generalization ability evaluation.

[0072] In a preferred embodiment of the present invention, the specific data collection and preprocessing method in step 1 includes:

[0073] We collected 2011 samples from five datasets: South Lifespan, Citigroup Biomedical Imaging Center (CBIC), Australian Imaging Biomarkers and Lifestyle Study of Ageing (AIBL), IXI (Information extraction from Images), and OASIS (Open Access Series of Imaging Studies).

[0074] All data were preprocessed using a unified pipeline, which mainly included grayscale correction of MRI structural brain images, skull stripping, and the application of nonlinear registration methods to align brain images to the MNI standard space.

[0075] Construct a data set and divide the preprocessed brain imaging data into training set, validation set, and test set in a ratio of 8:1:1.

[0076] In another preferred embodiment of the present invention, the step 2 specifically includes model construction to achieve brain age prediction. This model is based on a task learning method of adversarial learning to achieve brain age prediction:

[0077] Main Task Model Construction: The model consists of two main parts. The first is the encoder, which is primarily composed of six convolutional modules. Each convolutional module contains a convolutional layer, a normalization layer, a pooling layer, and an activation layer. The number of channels corresponding to the six convolutional modules is 32, 64, 128, 256, 256, and 128, respectively. The second part is the module for the brain age prediction task. Brain age prediction is mainly divided into two parts. The first is a traditional regression task, which predicts brain age based on the final linear activation layer and estimates the results based on the L1 loss function. The second part is to address the uneven age distribution of samples during training. Since most samples are not evenly distributed across age groups, a large proportion are concentrated in a certain age range. Directly predicting brain age can result in large errors during regression. Therefore, to optimize this problem, a specific age is mapped to the entire range according to the normal distribution. In this case, the range is 5-94. The mapping formula is as follows:

[0078]

[0079] Where y is the corresponding age label, μ and σ correspond to the expectation and variance of the normal distribution. In the experiment, we take the normal distribution with an expectation of 0 and a mean of 1. Therefore, for each hashed label, the KL divergence is used to constrain the output of the result. Therefore, the total loss function of this part of the task is:

[0080]

[0081] where y i It represents the true brain age. represents the estimated brain age, Represents y i The hashed tag value;

[0082] Adversarial learning task model building: In this step, the ultimate goal is to achieve accurate estimation of brain age. Therefore, in this step, the encoder is reduced to extracting features related to gender and site information, so that the encoder can extract features related to age as much as possible. Therefore, in this part, we set up multiple mixing factor modules to make the features extracted by the encoder independent of these mixing factors (site, gender). First, we implement the classification task of multiple mixing factors, that is, use cross entropy to constrain the classification mixing factor task. The main loss function form is as follows:

[0083]

[0084] Where i is the number of mixing factor modules, C i Represents the output category of factor classification, d c represents the output of the classifier;

[0085] Based on the common loss function task, minimizing the loss function means that the encoder is maximized in extracting the features related to the task. Therefore, our goal is to minimize the encoder's extraction of features related to the mixing factor. Therefore, we add a gradient reversal layer between the encoder and the mixing factor module to minimize the encoder's extraction of the mixing factor features. Therefore, the total loss function is:

[0086] Loss=Loss age +λLoss c_f

[0087] Where λ corresponds to the weight of the loss function of the mixing factor, here we take 0.8;

[0088] According to the principle of domain adaptation, after the gradient is updated by the mixed factor model, E is realized. si (encoder mapping on gender domain) and E di There is no significant difference in the (encoder mapping between site domains) training samples;

[0089] Adversarial learning training steps: To achieve the final task of brain age prediction and minimize the weights of the feature regions related to the mixing factor, we set a hyperparameter ε to balance the adversarial task. The main training steps are as follows:

[0090] Input: Training set D = {(x(n), y(n))}, validation set V, learning rate α, hyperparameter ε, regularization coefficient β of the loss function, encoder E i , mixing factor module M i , brain age prediction module BA i , adversarial task parameter con_st = 0.5;

[0091] Randomly initialize the network weights w i , b i ε = 0 (w i ∈ w, b i ∈ B)

[0092] Train and optimize through the neural network model to obtain (w, b);

[0093] The specific training process is as follows:

[0094] Set the total number of loops K;

[0095] Randomly reorder the samples in the training set D, divide the samples in D into N batches, and the number of samples in each batch is 8;

[0096] For the samples in the L-th batch, update the parameter ε = (i + L * i) / (N * k), L < N, and i is the current number of loops;

[0097] Select samples (x i , y i ) from the training set D, calculate the outputs of the image features passing through E i , M i , BA i and use the loss function to constrain the backpropagation calculation of the error δ of the gradient passed back by the mixing factor d , the gradient error δ of brain age prediction BA , to optimize the parameters of the encoder's gradient. The error of the reverse gradient is:

[0098]

[0099] Reverse-calculate the derivatives of the parameters of each layer of the encoder according to the loss function;

[0100] Update the parameters;

[0101] Perform the i+1th loop until the error rate of the neural network model on the validation set V no longer decreases, and then output (W, B), where W represents the weights of all neurons in the neural network and B represents the bias.

[0102] At the beginning of training, ∈ is approximately 0, so the encoder learns only features related to the brain age label. As epochs increase, the value of ∈ increases. That is, during training, the gradient coefficient of the gradient propagation of the mixing factor becomes increasingly larger, forcing the encoder to minimize brain image features related to the mixing factor. Through continuous adversarial learning, the encoder maximizes the extraction of features related to brain age while significantly reducing the error caused by site effects and gender differences in the model.

[0103] In another preferred embodiment of the present invention, the model performance evaluation and model generalization exploration in step 3 specifically include:

[0104] (1) Model prediction accuracy index: The model of this patent is mainly used to accurately estimate brain age. Therefore, in order to measure the prediction performance of brain age, the accuracy of brain age prediction is quantified by three indicators:

[0105] 1. Mean absolute error, as follows:

[0106]

[0107] Where y is the true label, Represents the predicted label. The lower the MAE, the higher the predicted age accuracy.

[0108] 2. Correlation coefficient r

[0109]

[0110] in represents the mean of the predicted values, Represents the mean of the true values, used to evaluate the similarity between the distribution of true and predicted values. The closer r is to 1, the better the prediction result. A value greater than or equal to 0.8 indicates that there is a high similarity between the predicted brain age and the true age, indicating a good prediction effect.

[0111] 3. R ​​Squared error

[0112]

[0113] in Represents the average value of the output, which is used to evaluate the effect of the model fitting the data. The closer it is to 1, the better the model fitting effect. 2If it is too low, it indicates that the model fitting effect is very poor and the model does not have the ability to accurately predict the data distribution.

[0114] (2) Brain age adaptive correction: For the prediction of brain age output by the model, due to the individual differences of the samples, we set the threshold of the output brain age prediction (Th max =7, the difference between the actual brain age and the predicted brain age), and correcting samples whose predicted brain age exceeds the threshold to increase the robustness and accuracy of the results. The main brain age correction method is based on a linear fitting model, as follows:

[0115] p a =α×Ω+β

[0116] Where Ω is the real age, p a Represents the predicted brain age. On the test set, the predicted brain age and the actual age data are used to fit α and β. Then, the predicted brain age is corrected by the difference using the following two methods:

[0117] C a =p a +[Ω-(α×Ω+β)]

[0118] Among them C a It is the corrected age.

[0119] (3) Generalization ability index: The generalization ability index is mainly divided into two aspects. The first is the brain age prediction accuracy of the main task pre-training model on other data sets. The second is to evaluate the shielding performance of the feature extractor on the mixed factor features, that is, to ensure that the feature extractor only extracts features related to brain age. For the evaluation of this result, the performance is evaluated by observing the mapping of the features of the autoencoder in the high-dimensional space and by observing the distribution in the high-dimensional space.

[0120] (4) Performance comparison method: The method proposed in this invention mainly realizes a more accurate prediction of brain age by adding a mixed factor module, and greatly improves the stability and generalization ability of the model. Therefore, in order to compare the reliability of the model, we compared conventional machine learning methods and more popular deep learning methods. According to previous studies, the features of the whole brain have a better effect on predicting brain age than the model that only uses the features of gray matter. Therefore, the deep learning method mainly predicts brain age by inputting the features of the whole brain into the model. To maintain consistency, in machine learning, we extracted all the features of the gray matter and white matter of the whole brain, performed feature dimensionality reduction through PCA, and used them as the input of the model to predict brain age. The compared methods mainly include two types, one is to predict brain age based on linear regression (lasso, elastic network, ridge), and the other is to predict brain age based on kernel methods, mainly including support vector regression and Gaussian process regression. The comparison method with deep neural networks is mainly based on the relatively conventional Resnet network architecture and the deep learning model based on the simple neural network structure (SFCN). Currently, these two methods are widely used in brain age model applications. The specific comparison results of the models are in Figure 3 visible.

[0121] The above is a detailed description of an embodiment of the present invention. However, the content described is only a preferred embodiment of the present invention and should not be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A brain age estimation method based on adversarial learning, characterized in that: Based on the whole-brain T1 magnetic resonance imaging, the corresponding age-specific imaging features are extracted to predict brain age. The following steps are involved: Step 1: Data collection and preprocessing: Collect data from more than 2,000 subjects across five sites, ranging in age from 5 to 94 years old. Preprocess the collected data and align all brain imaging data to the MNI standard space to standardize the size. Step 2: Based on the adversarial learning deep neural network, the network is trained and optimized to predict brain age; Step 3: Evaluate the performance of the brain age prediction task, optimize the prediction results, and evaluate the model's feature extraction and generalization capabilities; In the step 2, the method for establishing the adversarial learning task model includes the following steps: To achieve the classification task of multiple mixed factors, that is, to use cross entropy to constrain the classification mixed factor task, the main loss function form is as follows: in, is the number of mixing factor modules, represents the output category of factor classification, represents the output of the classifier; Based on the common loss function task, the loss function is minimized, that is, the encoder is allowed to extract the features related to the task to the maximum extent, and the encoder is allowed to extract the features related to the mixing factor to the minimum extent. A gradient reversal layer is added between the encoder and the mixing factor module to minimize the encoder's extraction of the mixing factor features. Therefore, the total loss function is: in It corresponds to the weight of the loss function of the mixing factor, and here we take 0.

8.

2. The brain age estimation method based on adversarial learning according to claim 1, characterized in that The specific data collection and preprocessing methods in step 1 include: We collected 2011 samples from five datasets: South lifespan, CBIC, AIBL, IXI, and OASIS; All data were preprocessed using a unified pipeline. This included grayscale correction of the MRI structural brain images, skull stripping, and registration of the brain images to the MNI standard space using a nonlinear registration method. Establish a data set and divide the preprocessed brain imaging data into training set, validation set, and test set in a ratio of 8:1:

1.

3. The brain age estimation method based on adversarial learning according to claim 1, characterized in that In step 2, the adversarial learning deep neural network includes a main task model and an adversarial learning task model, wherein the main task model includes two parts, namely an encoder part and a brain age prediction part. The method for building the main task model specifically includes the following steps: The first part is the encoder, which consists of 6 convolutional modules. Each convolutional module contains a convolution layer, a normalization layer, a pooling layer, and an activation layer. The number of channels corresponding to the 6 convolutional modules is 32, 64, 128, 256, 256, and 128 respectively. The second part is the brain age prediction part. The brain age prediction includes two parts: sample optimization and regression calculation. The brain age is predicted based on the last linear activation layer, and the result is estimated based on the L1 loss function.

4. The method for estimating brain age based on adversarial learning according to claim 3, characterized in that: In step 2, the specific steps of sample optimization in the brain age prediction part are as follows: The corresponding age is mapped to the range of 5-94 with a normal distribution of expected value of 0 and mean value of 1. The mapping formula is as follows: Where y is the label of the corresponding age, , Corresponding to the expectation and variance of the normal distribution, for each hashed label, the KL divergence is used to constrain the output of the result. The total loss function of this part of the task is: in It represents the true brain age. represents the estimated brain age, represent The hashed tag value.

5. The brain age estimation method based on adversarial learning according to claim 1, characterized in that: The specific training steps in step 2 are as follows: Input: training set , validation set , learning rate , hyperparameters , the regularization coefficient of the loss function , encoder , Mixing Factor Module ,Brain Age Prediction Module , adversarial task parameter con_st=0.5; Randomly initialize network weights , ; Through training and optimization of the neural network model, we can obtain ; The specific training process is as follows: Set the total number of loops K; For the training set D The samples in are randomly reordered, D The sample is divided into batches, the number of samples in each batch is 8; For the first samples in batches, update parameters , , is the current cycle number; From the training set Select sample ( , ), calculate the image features through , , The output of , and the loss function is used to constrain back propagation to calculate the error of the gradient returned by the mixing factor , gradient error of brain age prediction , to optimize the parameters of the encoder gradient, the error of the reverse gradient is: ； Inversely calculate the derivatives of the parameters of each layer of the encoder according to the loss function; Update parameters; Carry out the +1 cycle, until the neural network model is in the validation set Output when the error rate no longer decreases ,W represents the weights of all neurons in the neural network, and B represents the bias.

6. The method for estimating brain age based on adversarial learning according to claim 1, characterized in that: In step 3, the specific method for evaluating the performance of the brain age prediction task is as follows: Calculate the mean absolute error: ;in, is the true label, Represents the predicted label. The lower the MAE, the higher the predicted age accuracy. Calculate the correlation coefficient r: in, represents the mean of the predicted values, Represents the mean of the true value, which is used to evaluate the similarity of the distribution between the true value and the predicted value. r The closer it is to 1, the better the prediction result; greater than or equal to 0.8 means that there is a greater similarity between the predicted brain age and the actual age, and the prediction effect is good; Calculate R square error: in, It represents the average value of the output, which is used to evaluate the effect of the model fitting the data. The closer it is to 1, the better the model fitting effect. The lower it is, the worse the model fitting effect is.

7. The method for estimating brain age based on adversarial learning according to claim 6, characterized in that: The step 3 also includes brain age adaptive correction, by setting the output brain age prediction threshold , to correct samples whose predicted brain age exceeds a threshold. The threshold for brain age prediction is the difference between the actual brain age and the predicted brain age. The specific steps of the brain age correction method are as follows: Establish a linear fitting model: in It's the real age. represents the predicted brain age, and The predicted brain age and the actual age data are used to fit the test set; Correction by interpolation: = +[ ] in It is the corrected brain age.

Images