Brain age prediction method based on deep learning and noise label learning

Abstract

The invention discloses a brain age prediction method based on deep learning and noise label learning, and belongs to the technical field of medical image processing. Aiming at the problem that the physiological age is possibly inconsistent with the actual brain age and the real brain age can not be accurately predicted by a training model due to the fact that the prediction is carried out by completely depending on the physiological age, the method combines two deep neural networks and a corresponding mean value teacher network; based on the consistency and difference between the brain age and the physiological age, clean label samples and noise label samples are identified and utilized, so that the prediction accuracy is improved. The method is excellent in performance on an analog data set containing noise; experimental results on a real data set show that the correlation between the cognitive index and the predicted brain age is higher than the correlation between the cognitive index and the physiological age. Therefore, the prediction method provided by the invention can more accurately capture the biological characteristics of brain aging.

Description

Technical Field

[0001] The present invention belongs to the field of medical image processing, and specifically relates to a brain age prediction method combining deep learning and noise label learning. Background Art

[0002] At present, with the increasing prominence of the aging population, the importance of accurately predicting brain age has become more and more significant. Studies have shown that there may be differences between brain age and physiological age. Although many studies have used brain imaging data to predict brain age, most studies still use physiological age as a label to train models, which may result in predictions that do not truly reflect the actual age of the brain. The lack of real brain age labels makes it a major challenge to accurately build brain age prediction models based on brain images. This is because the model is forced to map neurobiological patterns to physiological age labels that may not fully reflect the true brain age. Although some studies have improved the accuracy of age prediction through bias correction algorithms or correlation-constrained regression methods, these methods are too simple, resulting in the generated models may not accurately predict the true brain age.

[0003] Therefore, there is an urgent need to develop an adaptive brain age prediction method driven by brain imaging data to reduce the negative impact of samples with large differences between physiological age and brain age on model training, thereby achieving more accurate brain age prediction. Summary of the invention

[0004] In view of the inconsistency between physiological age and actual brain age, relying solely on physiological age for prediction may result in the training model being unable to accurately predict the true brain age. The present invention proposes a brain age prediction method that combines deep learning with noise label learning, which can effectively achieve accurate prediction of brain age. This method effectively identifies and fully utilizes the brain image features and label information of subjects whose brain age is consistent with their physiological age, as well as the brain image feature information of subjects whose brain age is inconsistent with their physiological age, thereby training a more accurate and robust brain age prediction model.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] The brain age prediction method based on deep learning and noisy label learning includes the following steps:

[0007] Step 1: construct a training data set;

[0008] Step 1.1, first preprocess the brain image data of the subject to obtain the standard brain image of the subject, use the brain image analysis method to estimate the brain image measurement from the brain image data, and extract the brain image features;

[0009] Step 1.2: Construct the subject's brain image feature matrix

[0010] Where N represents the number of subjects, D represents the number of brain imaging features of the subjects; x in the matrix X i represents the brain image features of the i-th subject, i∈{1,2,…,N};

[0011] Step 1.3: Construct the physiological age label vector corresponding to the subject

[0012] Where N represents the number of subjects, is the i-th row of the Y vector, representing the physiological age of the i-th subject, i∈{1,2,…,N};

[0013] Step 1.4, set two empty subject subscript sets S Clean and S Noisy ;

[0014] Among them, S Clean The subscript used to store the subjects whose brain age and physiological age are relatively consistent identified by this method, S Noisy Subscripts used to store subjects whose brain age and biological age are not consistent enough;

[0015] Step 2: training the model;

[0016] Step 2.1, initialize two main deep neural networks θ1 and θ2, and two mean teacher neural networks corresponding to the two main deep neural networks θ1 and θ2 and

[0017] Step 2.2: Input the brain image feature matrix X and the corresponding physiological age label vector Y in the training data set into the two main deep neural networks θ1 and θ2 of the model; the main deep neural networks θ1 and θ2 output the prediction results for each subject respectively. and

[0018] Specifically, for each subject i, where i∈{1,2,…,N}, its similarity consistency loss is calculated To evaluate the similarity between the subject's predicted value and the original label and the consistency between the prediction results of the two main networks, the formula is as follows:

[0019]

[0020] Among them, λ1 is a weight parameter used to balance the similarity between the predicted value and the original label and the consistency between the prediction results of the two main deep neural networks;

[0021] Step 2.3: Substitute the similarity consistency loss term for each subject in the training dataset Sort them in ascending order; if the similarity consistency loss value of the subject is lower than the set threshold, it is identified as a clean sample, and its subscript i is recorded in the subject subscript set S whose brain age and physiological age are relatively consistent Clean On the contrary, if the similarity consistency loss value of the subject is higher than the set threshold, it is identified as a noise sample, and its subscript i is recorded in the subscript set S of the subject whose brain age and physiological age are not consistent Noisy middle;

[0022] Step 2.4, input the brain image features of each subject into two mean teacher neural networks and The prediction results are obtained and

[0023] Step 2.5, for each subject i, where i∈{1,2,…,N}, calculate the teacher loss To evaluate the similarity between the predictions of the main deep neural network and the mean teacher network, the formula is as follows:

[0024]

[0025] For each subject i whose brain age and biological age are relatively consistent, where i∈S Clean , and its similarity consistency loss Loss of Teachers Sum the total loss for each subject whose brain age and physiological age are relatively consistent; then sum the total losses of all subjects whose brain age and physiological age are relatively consistent to calculate the overall clean loss term |S Clean | indicates S Clean The number of subject subscripts whose brain age and biological age stored in are relatively consistent;

[0026]

[0027] Compute the prediction consistency loss for each subject i whose brain age is inconsistent with biological age

[0028]

[0029] For each subject i whose brain age is inconsistent with his biological age, where i∈S Noisy , and its prediction consistency loss term Loss of Teachers Sum the total loss for each subject whose brain age and physiological age are inconsistent; then sum the total losses of all subjects whose brain age and physiological age are inconsistent to calculate the overall noise loss term |S Noisy | indicates S Noisy The brain age and biological age stored in the data are not consistent enough. The number of subjects' subscripts;

[0030]

[0031] Step 2.6, calculate the sum of losses for all subjects

[0032]

[0033] Among them, λ2 is a weight parameter used to balance the loss of subjects with relatively consistent brain age and physiological age. Subject loss not consistent with brain age and biological age

[0034] Step 2.7, update the parameters of the main deep neural network θ1 and θ2 using the gradient descent method, and update the mean teacher network by the exponential moving average method and Parameters; if the total loss If converged, complete the training and save the model; if not, return to step 2.2 and continue training until converged;

[0035] Step 3: Extract brain image features from the new sample data, input them into the trained model for analysis, output the prediction results, and obtain the estimated brain age.

[0036] Furthermore, the preprocessing in step 1 specifically includes inter-layer correction, head motion correction, spatial normalization and spatial smoothing.

[0037] Compared with the prior art, the present invention has the following advantages:

[0038] 1) The present invention proposes a brain age prediction method based on deep learning and noisy label learning, which can accurately predict brain age and overcome the limitation of existing methods that rely on physiological age labels. The present invention predicts brain age through brain imaging data, effectively identifies and utilizes clean and noisy label samples, thereby improving the accuracy of prediction.

[0039] 2) The present invention not only uses the brain image features and labels of subjects whose brain age and physiological age are relatively consistent to explore the mapping between brain images and labels, but also effectively uses the brain image features of subjects whose brain age and physiological age are inconsistent to assist model learning. This method can improve the robustness of model prediction in the presence of label noise.

[0040] 3) The present invention reduces the negative impact of samples with inconsistent brain age and physiological age on training, and trains a more accurate and reliable brain age prediction model. This method helps doctors predict brain age more accurately and provides new ideas for neuroscience and brain aging research. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a flowchart of a method for predicting brain age based on deep learning and noise label learning according to the present invention;

[0042] Figure 2 This is a schematic diagram of the test set results of the model predicting brain age under different label noise rates in the simulated data set;

[0043] Figure 3 This is a schematic diagram of the correlation results between cognitive indicators and physiological age and predicted age. DETAILED DESCRIPTION

[0044] In order to gain a deeper understanding of the present invention, we will provide a comprehensive and detailed description of the present invention. However, the present invention has multiple implementations and is not limited to the specific examples listed herein. The presentation of these examples is intended to deepen the comprehensive understanding of the disclosure of the present invention.

[0045] The brain age prediction method based on deep learning and noisy label learning includes the following steps:

[0046] Step 1: construct a training data set;

[0047] Step 1.1, brain image data from normal subjects are collected from the site, and preprocessing operations including inter-layer correction, head motion correction, spatial normalization and spatial smoothing are performed on these data to obtain standard brain images of the subjects, and brain image analysis methods are used to estimate brain image measurements from the brain image data, and feature selection or feature extraction methods are used to extract features from the brain image measurements;

[0048] Step 1.2: Construct the subject's brain image feature matrix

[0049] Where N represents the number of subjects, D represents the number of brain imaging features of the subjects; x in the matrix X i represents the brain image features of the i-th subject, i∈{1,2,…,N};

[0050] Step 1.3: Construct the physiological age label vector corresponding to the subject

[0051] Where N represents the number of subjects, is the i-th row of the Y vector, representing the physiological age of the i-th subject, i∈{1,2,…,N};

[0052] Step 1.4, set two empty subject subscript sets S Clean and S Noisy ;

[0053] Among them, S Clean The subscript used to store the subjects whose brain age and physiological age are relatively consistent identified by this method, S Noisy Subscripts used to store subjects whose brain age and biological age are not consistent enough;

[0054] Step 2: training the model;

[0055] Step 2.1, initialize two main deep neural networks θ1 and θ2, and two mean teacher neural networks corresponding to the two main deep neural networks θ1 and θ2 and

[0056] Step 2.2: Input the brain image feature matrix X and the corresponding physiological age label vector Y in the training data set into the two main deep neural networks θ1 and θ2 of the model; the main deep neural networks θ1 and θ2 output the prediction results for each subject respectively. and

[0057] Specifically, for each subject i, where i∈{1,2,…,N}, its similarity consistency loss is calculated To evaluate the similarity between the subject's predicted value and the original label and the consistency between the prediction results of the two main networks, the formula is as follows:

[0058]

[0059] Among them, λ1 is a weight parameter used to balance the similarity between the predicted value and the original label and the consistency between the prediction results of the two main deep neural networks;

[0060] Step 2.3: Substitute the similarity consistency loss term for each subject in the training dataset Sort them in ascending order; if the similarity consistency loss value of the subject is lower than the set threshold, it is identified as a clean sample, and its subscript i is recorded in the subject subscript set S whose brain age and physiological age are relatively consistent Clean On the contrary, if the similarity consistency loss value of the subject is higher than the set threshold, it is identified as a noise sample, and its subscript i is recorded in the subscript set S of the subject whose brain age and physiological age are not consistent Noisy middle;

[0061] Step 2.4, input the brain image features of each subject into two mean teacher neural networks and The prediction results are obtained and

[0062] Step 2.5, for each subject i, where i∈{1,2,…,N}, calculate the teacher loss To evaluate the similarity between the predictions of the main deep neural network and the mean teacher network, the formula is as follows:

[0063]

[0064] For each subject i whose brain age and biological age are relatively consistent, where i∈S Clean , and its similarity consistency loss Loss of Teachers Sum the total loss for each subject whose brain age and physiological age are relatively consistent; then sum the total losses of all subjects whose brain age and physiological age are relatively consistent to calculate the overall clean loss term |S Clean | indicates S Clean The number of subject subscripts whose brain age and biological age stored in are relatively consistent;

[0065]

[0066] Compute the prediction consistency loss for each subject i whose brain age is inconsistent with biological age

[0067]

[0068] For each subject i whose brain age is inconsistent with his biological age, where i∈S Noisy , and its prediction consistency loss term Loss of Teachers Sum the total loss for each subject whose brain age and physiological age are inconsistent; then sum the total losses of all subjects whose brain age and physiological age are inconsistent to calculate the overall noise loss term |S Noisy | indicates S Noisy The brain age and biological age stored in the data are not consistent enough. The number of subjects' subscripts;

[0069]

[0070] Step 2.6, calculate the sum of losses for all subjects

[0071]

[0072] Among them, λ2 is a weight parameter used to balance the loss of subjects with relatively consistent brain age and physiological age. Subject loss not consistent with brain age and biological age

[0073] Step 2.7, update the parameters of the main deep neural network θ1 and θ2 using the gradient descent method, and update the mean teacher network by the exponential moving average method and Parameters; if the total loss If converged, complete the training and save the model; if not, return to step 2.2 and continue training until converged;

[0074] Step 3: Extract brain image features from the new sample data, input them into the trained model for analysis, output the prediction results, and obtain the estimated brain age.

[0075] Experiments were conducted on simulated data with different degrees of label noise. First, a sample size of 10,000 with a brain age label Y ranging from approximately 45 to 75 years old was used. b . We defined 10 potential factors for variation in brain imaging measures, where the first factor represents brain age and the remaining nine factors account for random variation. These factors were combined using a sparse mixing matrix of dimension 10×3000, which maps each factor to multiple imaging variables, creating a set of 3000 imaging variables for each sample. This resulted in a feature matrix X of dimension 10000×3000. All columns in X were then standardized to form a standardized feature matrix. Then, the brain age labels (Y b ) Gaussian noise with different noise rates (20%, 40%, 80%) was applied to generate noisy labels Y. Under the influence of noise, Y in the training set and brain age label Y b The Pearson correlation coefficients between them are 0.849, 0.749, and 0.623, respectively. For the test set samples without noise, the Y value is equal to Y b The model was then trained using the training set, and the prediction accuracy of the model was evaluated on the test set. Figure 2 The scatter plot of predicted brain age and real brain age under different noise rates is shown. It can be seen from the figure that under different label noise levels, the predicted age of the test data (predicted label) and the simulated brain age (real label) are highly consistent, with a correlation of more than 0.9, and the prediction results of the two main networks remain highly consistent, indicating that under different noise conditions, our model can maintain a small prediction error and high prediction accuracy.

[0076] At the same time, the present invention also conducted experiments on brain imaging data of healthy subjects in the real data set UKBiobank. The effect of the method was evaluated by calculating the correlation between cognitive indicators and predicted brain age and physiological age. Figure 3 The correlations between cognitive indicators and predicted brain age and physiological age are shown. It can be seen from the figure that the correlation between the model-predicted brain age and cognitive measures exceeds the correlation between physiological age and cognitive measures, indicating that the predicted brain age produced by our method is closer to the biological aspects of aging; Figure 3 The cognitive indicators in the test include fluid intelligence (FI), numeric memory (NM) and reaction time (RT).

[0077] The contents not described in detail in the specification of the present invention belong to the prior art known to the professional and technical personnel in the field. Although the illustrative specific embodiments of the present invention are described above to facilitate the understanding of the present invention by the technical personnel in the field, it should be clear that the present invention is not limited to the scope of the specific embodiments. For the ordinary technical personnel in the field, as long as various changes are within the spirit and scope of the present invention defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.

Claims

1. A brain age prediction method based on deep learning and noisy label learning, characterized in that: Step 1, construct a training data set for the noisy label learning model; Step 1.1, first preprocess the brain image data of the subject to obtain the standard brain image of the subject, use the brain image analysis method to estimate the brain image measurement from the brain image data, and extract the brain image features; Step 1.2: Construct the subject's brain image feature matrix Where N represents the number of subjects, D represents the number of brain imaging features of the subjects; x in the matrix X i represents the brain image features of the i-th subject, i∈{1,2,…,N}; Step 1.3: Construct the physiological age label vector corresponding to the subject Where N represents the number of subjects, is the i-th row of the Y vector, representing the physiological age of the i-th subject, i∈{1,2,…,N}; Step 1.4, set two empty subject subscript sets S Clean and S Noisy ; Among them, S Clean The subscript used to store the subjects whose brain age and physiological age are relatively consistent identified by this method, S Noisy Subscripts used to store subjects whose brain age and biological age are not consistent enough; Step 2: training the model; Step 2.1, initialize two main deep neural networks θ1 and θ2, and two mean teacher neural networks corresponding to the two main deep neural networks θ1 and θ2 and Step 2.2: Input the brain image feature matrix X and the corresponding physiological age label vector Y in the training data set into the two main deep neural networks θ1 and θ2 of the model; the main deep neural networks θ1 and θ2 output the prediction results for each subject respectively. and i∈{1,2,…,N}; Specifically, for each subject i, where i∈{1,2,…,N}, its similarity consistency loss is calculated To evaluate the similarity between the subject's predicted value and the original label and the consistency between the prediction results of the two main networks, the formula is as follows: Among them, λ1 is a weight parameter used to balance the similarity between the predicted value and the original label and the consistency between the prediction results of the two main deep neural networks; Step 2.3: Substitute the similarity consistency loss term for each subject in the training dataset i∈{1,2,…,N} is sorted in ascending order; if the similarity consistency loss value of the subject is lower than the set threshold, it is identified as a clean sample, and its subscript i is recorded in the subject subscript set S whose brain age and physiological age are relatively consistent Clean On the contrary, if the similarity consistency loss value of the subject is higher than the set threshold, it is identified as a noise sample, and its subscript i is recorded in the subscript set S of the subject whose brain age and physiological age are not consistent Noisy middle; Step 2.4, input the brain image features of each subject into two mean teacher neural networks and The prediction results are obtained and i∈{1,2,…,N}; Step 2.5, for each subject i, where i∈{1,2,…,N}, calculate the teacher loss To evaluate the similarity between the predictions of the main deep neural network and the mean teacher network, the formula is as follows: For each subject i whose brain age and biological age are relatively consistent, where i∈S Clean , and its similarity consistency loss Loss of Teachers Sum the total loss for each subject whose brain age and physiological age are relatively consistent; then sum the total losses of all subjects whose brain age and physiological age are relatively consistent to calculate the overall clean loss term |S Clean | indicates S Clean The number of subject subscripts whose brain age and biological age stored in are relatively consistent; Compute the prediction consistency loss for each subject i whose brain age is inconsistent with biological age i∈S Noisy ; For each subject i whose brain age is inconsistent with his biological age, where i∈S Noisy , and its prediction consistency loss term Loss of Teachers Sum the total loss for each subject whose brain age and physiological age are inconsistent; then sum the total losses of all subjects whose brain age and physiological age are inconsistent to calculate the overall noise loss term |S Noisy | indicates S Noisy The brain age and biological age stored in the data are not consistent enough. The number of subjects' subscripts; Step 2.6, calculate the sum of losses for all subjects Among them, λ2 is a weight parameter used to balance the loss of subjects with relatively consistent brain age and physiological age. Subject loss not consistent with brain age and biological age Step 2.7, update the parameters of the main deep neural network θ1 and θ2 using the gradient descent method, and update the mean teacher network by the exponential moving average method and Parameters; if the total loss If converged, complete the training and save the model; if not, return to step 2.2 and continue training until converged; Step 3: Extract brain image features from the new sample data, input them into the trained model for analysis, output the prediction results, and obtain the estimated brain age.

2. The accurate brain age prediction method based on noisy label learning according to claim 1 is characterized in that: The preprocessing in step 1 specifically includes inter-layer correction, head motion correction, spatial normalization and spatial smoothing.

Images