A method for predicting brain age based on automatic fiber bundle recognition

Abstract

A method for predicting brain age based on automatic fiber bundle recognition. Aiming at the problems existing in the existing method for predicting brain age by extracting white matter features based on voxels, such as the inability to accurately locate fiber bundles sensitive to age changes and the averaging of fiber characteristics, the diffusion tensor imaging images are used to obtain a whole-brain fiber bundle map through a fiber tracking algorithm. Anatomically significant fiber bundles are automatically segmented and recognized according to the brain white matter atlas, and then diffusion metrics are quantified along the fiber bundles. Finally, the diffusion metrics are used as features to input into the brain age prediction model to test the accuracy of the prediction model. The present invention extracts tensor values as white matter features through the method of automatic fiber bundle recognition, and constructing a brain age prediction model can better reflect the subtle changes in the microstructure of white matter fiber bundles during the aging process and find out the fiber bundles sensitive to changes during the aging process.

Description

Technical Field

[0001] The present invention relates to the fields of medical image processing and machine learning, and particularly to a brain age prediction method based on automatic clustering and annotation of fiber bundles. Background Art

[0002] As a non-invasive in-vivo method, magnetic resonance imaging (MRI) technology has been widely used in neuroscience research, and it has played a great role in obtaining clinical nerve fiber structure information and assisting in understanding the functional connections between different regions of the brain. The brain has very good plasticity. During the entire human life cycle, the brain structure will change to a certain extent due to factors such as physiological development, cognitive activities, and diseases. The brain age prediction framework enables us to evaluate an individual's brain health status. If the predicted brain age is within the predicted normal change range, it can be considered that the individual's brain is healthy. Existing studies have applied the brain age prediction framework to patients with various neurological diseases, such as Alzheimer's disease, mild cognitive impairment, traumatic brain injury, and intractable epilepsy, and found that the predicted brain age is significantly older than the actual age. Therefore, the difference between the predicted brain age and the actual age is considered a potential imaging biomarker for judging whether the brain is healthy, which can identify brain degeneration and may contribute to the early diagnosis of neurodegenerative diseases.

[0003] Most current brain age prediction models extract features of gray matter or the entire brain structure, such as cortical volume and cortical thickness, from T1-weighted images. However, existing studies have shown that compared with gray matter, the microstructure of white matter fiber bundles may be more sensitive to subtle changes in the aging brain. Although the white matter microstructure has potential value for brain age prediction, only a few studies have attempted to extract white matter features from magnetic resonance diffusion tensor imaging (DTI) data to predict an individual's brain age. Moreover, existing research methods all extract fractional anisotropy (FA), mean diffusivity (MD), radial diffusivity (RD), and axial diffusivity (AD) as white matter features based on the voxel level. This voxel-level method can only calculate the mean value of the region of interest, which is prone to averaging fiber characteristics and cannot accurately locate the white matter fiber bundles that are more sensitive to brain age. Summary of the Invention

[0004] In order to overcome the deficiencies of existing methods for predicting brain age by extracting white matter features based on voxels, which are unable to accurately locate fiber bundles sensitive to age changes and will cause the averaging of fiber characteristics, the present invention proposes a method for predicting brain age based on automatic fiber bundle recognition. Specifically, the present invention first uses a fiber tracking method to obtain a whole-brain fiber bundle map, automatically recognizes the fiber bundles using the method of the brain white matter atlas, then uses the BUAN algorithm to perform position-based sampling and quantization along a bundle of fibers, and the white matter features obtained thereby are trained to obtain a brain age prediction model after feature screening, and this model is applied to test data to evaluate the generalization ability of the model.

[0005] The technical solution adopted by the present invention to solve its technical problems is as follows:

[0006] A method for predicting brain age based on automatic fiber bundle recognition, the method comprising the following steps:

[0007] Step 1: Perform image preprocessing on magnetic resonance diffusion tensor imaging, and the process is as follows:

[0008] Use FSL tools to perform noise reduction, head motion correction, eddy current correction, distortion correction, and EPI correction operations on the image;

[0009] Step 2: Trace the whole-brain fiber bundles, and the process is as follows:

[0010] Use the UKF fiber tracking method of double tensors to trace the whole-brain fiber bundles to obtain a whole-brain fiber bundle map;

[0011] Step 3: Automatically segment and recognize the fiber bundles of the whole brain, and the process is as follows:

[0012] Automatically segment the whole-brain fiber bundle map obtained in Step 2 using the brain white matter atlas. This method can segment small bundles of fibers that are not easily segmented manually, and obtain 13 bundles of fibers in each of the left and right brains and 6 bundles of connecting fibers;

[0013] Step 4: Extract tensor values FA, MD, RD, and AD along the fiber bundles as white matter features, and the process is as follows:

[0014] Use the BUAN algorithm to perform equidistant sampling based on position along the fiber bundle direction, sample the fiber bundle into a set number of segments, and calculate the average value of the diffusion tensor of each segment as the white matter feature;

[0015] Step 5: Perform feature selection on the features extracted in Step 4, train to obtain an age prediction model, and use the test set to test and evaluate the generalization ability of the model, and analyze the fiber bundles in the prediction model that are more sensitive to age, and the process is as follows:

[0016] Due to the problem of multicollinearity in the features extracted in Step 4, the features are first subjected to recursive feature elimination, and then the brain age prediction model is trained using ridge regression. The fiber bundles related to age changes are analyzed based on the model weights.

[0017] Furthermore, in the above-mentioned Step 3, the process of automatically identifying the whole-brain fiber bundle map is as follows:

[0018] First, the whole-brain fiber bundle map of each subject is registered into the brain white matter atlas space. According to the white matter partition generalization WMPG, the inter-subject partition feasibility ISPV, and the fiber bundle anatomical contour consistency TAPC, the whole-brain fiber bundles are clustered into 800 clusters. Each fiber in the whole-brain fiber bundle of each subject is assigned to the nearest atlas cluster in the atlas space. The fiber clusters of each subject are divided into left hemisphere bundles, right hemisphere bundles, and commissural bundles. After further removing abnormal fibers by comparing with the atlas bundles, the clusters with the same structure and function in the 800 clusters are identified as fiber bundles with anatomical significance, such as the superior longitudinal fasciculus SLF.

[0019] Even further, in the above-mentioned Step 4, the process of extracting white matter features along the fiber bundle is as follows:

[0020] First, the centroid of the atlas cluster in Step 3 is divided into 100 segments along the length in the atlas space to obtain a template cluster. Then, for the cluster to be segmented, the Euclidean distance between each point on each streamline of the cluster and the centroid of the template cluster is calculated, and the point is assigned to the nearest centroid segment;

[0021] The eigenvalues λ1, λ2, and λ3 in three directions in the three-dimensional space are obtained using the diffusion tensor matrix. The fractional anisotropy FA, mean diffusivity MD, radial diffusivity RD, and axial diffusivity AD of each point are calculated through the following calculations:

[0022]

[0023]

[0024]

[0025] AD = λ1

[0026] After each cluster is divided into 100 segments, the average values of the diffusion tensor values FA, MD, RD, and AD on each segment are calculated as the white matter features of the brain age prediction model.

[0027] Still further, in the above-mentioned Step 5, the training and testing process of the brain age prediction model is as follows:

[0028] 5.1) Feature Selection: In the first step of feature selection, the Pearson correlation coefficient between each feature in the fiber bundles with anatomical significance and the actual age is calculated. A positive Pearson correlation coefficient indicates a positive correlation between the feature and the actual age, while a negative coefficient indicates a negative correlation. The weights are calculated based on the correlation with age, and the weight calculation formula is as follows:

[0029]

[0030] where n represents the total number of fibers in a bundle of fibers, and r i represents the Pearson correlation coefficient between each white matter feature in the bundle and the actual age;

[0031] Then, the weighted average of the diffusion tensor values of each bundle of fibers is obtained. There are 33 bundles of fibers and 4 diffusion tensor values, so there are 33 * 4 = 132 eigenvalues. Since the problem of high feature dimensionality can lead to model overfitting, the recursive feature elimination method is used to try various feature combinations, and the feature combination with the best model training effect is selected as the model input;

[0032] 5.2) Feature Normalization: After dealing with feature outliers, the features are scaled so that the feature values are distributed between 0 and 1, and 80% of all samples are divided into the training set, with the remaining 20% as the test set;

[0033] 5.3) Training the Model: The feature combination obtained from feature selection is used as the eigenvalue, and the actual age of the training set samples is used as the label to input the ridge regression model for training. Due to the small amount of data, five-fold cross-validation is used, and the evaluation indicators of the five validation sets are averaged. The grid search is used to find the model parameters that make the model perform best. The coefficient of each term in this regression model is the weight of the feature, and the larger the coefficient, the more sensitive the bundle of fibers is to age;

[0034] 5.4) Testing the Model: The optimal feature combination of the training set is input into the model trained above to obtain the brain age predicted by the diffusion index. The evaluation indicators are calculated between the predicted brain age and the actual age, including the mean absolute error MAE, the root mean square error RMSE, and the Pearson correlation coefficient between the actual age and the predicted age, which are used to measure the accuracy and effectiveness of this brain age prediction model.

[0035] The beneficial effects of the present invention are as follows: It realizes the effective prediction of brain age using white matter features and can identify the fiber bundles that are sensitive to age changes during development and aging. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is a schematic flow chart of the implementation scheme of the present invention. DETAILED IMPLEMENTATION SCHEME

[0037] To make the objectives, technical solutions, and advantages of the present invention more clear and understandable, the following provides a further supplementary description of the present invention in conjunction with specific embodiments and drawings.

[0038] Referring to Figure 1 , a method for predicting brain age based on automatic fiber bundle recognition can effectively predict brain age through diffusion tensor imaging data, find fiber bundles sensitive to age, and judge the health status of the brain, including the following steps:

[0039] Step 1: Preprocessing of magnetic resonance diffusion tensor imaging images, the process is as follows:

[0040] During the magnetic resonance scanning of the subjects, there will be impacts on the image quality due to factors such as the subjects' breathing, head movement, and machine noise. Therefore, it is necessary to first use FSL tools to perform operations such as image denoising, head movement correction, eddy current correction, distortion correction, and EPI correction on the images to exclude potential artifacts caused by eddy currents, head movement, and magnetic field distortion.

[0041] Step 2: Tracing the whole-brain fiber bundles, the process is as follows:

[0042] The UKF fiber tracking method establishes a bi-tensor model and fits the input diffusion tensor signal to estimate the direction of fiber continuity. Voxels with FA greater than 0.15 are selected as masks, five initial seed points are set for each voxel, and the step size during the tracking process is set to 0.3 mm. Once the FA value at the current position is lower than 0.15 or the fiber trajectory reaches the boundary of the mask, the tracking is terminated. A whole-brain fiber bundle map containing approximately 350,000 fibers is generated for each subject.

[0043] Step 3: Automatically segmenting and identifying fiber bundles in the whole-brain fiber bundles, the process is as follows:

[0044] According to the brain white matter atlas, the whole-brain fiber bundle map obtained in Step 2 is automatically segmented into fiber bundles with anatomical significance. First, the whole-brain fiber bundle map of each subject is registered into the brain white matter atlas space, and the registration process includes non-rigid transformation and affine transformation; then, according to the white matter segmentation generalization WMPG, the inter-subject segmentation feasibility ISPV, and the fiber bundle anatomical contour consistency TAPC, the whole-brain fiber bundles are clustered into 800 clusters, and each fiber in the whole-brain fiber bundle of each subject is assigned to the nearest atlas cluster in the atlas space.

[0045] To improve the accuracy of segmentation, the bilateral clustering step is used to simultaneously segment the fibers into left and right hemisphere fibers and commissural fibers, and then it is determined again whether each fiber belongs to the corresponding cluster. If the probability of a fiber is more than two standard deviations higher than the average fiber probability of the cluster, the fiber is removed; finally, the clusters with the same structure and function in the 800 clusters are identified as fiber bundles with anatomical significance, including 13 fiber bundles in each of the left and right brains and 6 commissural fibers.

[0046] Step 4: Extract tensor values FA, MD, RD, and AD along the fiber bundles as white matter features. The process is as follows:

[0047] First, divide the centroid of the atlas cluster in step 3 into 100 segments along the length in the atlas space to obtain a template cluster. Then, for the cluster to be segmented, calculate the Euclidean distance between each point on each streamline of the cluster and the centroid of the template cluster, and assign the point to the nearest centroid segment;

[0048] Use the diffusion tensor matrix to obtain the eigenvalues λ1, λ2, and λ3 in three directions in three-dimensional space, and calculate the fractional anisotropy FA, mean diffusivity MD, radial diffusivity RD, and axial diffusivity AD of each point through the following calculations:

[0049]

[0050]

[0051]

[0052] AD = λ1

[0053] After dividing each cluster into 100 segments, calculate the average values of the diffusion tensor values FA, MD, RD, and AD on each segment as the white matter features of the brain age prediction model.

[0054] Step 5: Perform feature selection on the features extracted in step 4, train an age prediction model, and use the test set to test and evaluate the generalization ability of the model, and analyze the fiber bundles in the prediction model that are more sensitive to age. The process is as follows:

[0055] 5.1) Feature selection: The first step of feature selection is to first calculate the Pearson coefficient between each feature in the fiber bundles with anatomical significance and the actual age. A positive Pearson coefficient indicates that the feature is positively correlated with the actual age, and a negative Pearson coefficient indicates that the feature is negatively correlated with the actual age, and calculate the weight according to the correlation with age. The weight calculation formula is as follows:

[0056]

[0057] Among them, n represents the total number of fibers in a bundle of fibers, r iIndicates the Pearson coefficient of each white matter feature in the bundle with the actual age;

[0058] Then, the weighted average of the fiber diffusion tensor values of each bundle is obtained. There are 33 bundles of fibers and 4 diffusion tensor values, so there are 33 * 4 = 132 eigenvalues.

[0059] Since the problem of high dimensionality of features can lead to model overfitting, the recursive feature elimination method is used to try various feature combinations. The main idea of recursive feature elimination is to repeatedly construct a model, select the best features, set the selected features aside, and then repeat this process on the remaining features until all features are traversed. The order of elimination in this process is the ranking of the features. Finally, select the feature combination with the best model training effect as the model input;

[0060] 5.2) Feature normalization: After dealing with feature outliers, scale the features so that the feature values are distributed between 0 and 1, and divide 80% of all samples into the training set, and the remaining 20% is the test set;

[0061] 5.3) Train the model: Use the feature combination obtained by feature selection as the feature values, and the actual age of the training set samples as the labels to input the ridge regression model for training; Since the data volume is small, five-fold cross-validation is used, and the evaluation indicators of the five validation sets are averaged, and grid search is used to find the model parameters that make the model effect the best; The coefficient of each term in this regression model is the weight of the feature, and the larger the coefficient, the more sensitive the bundle of fibers is to age;

[0062] 5.4) Test the model: Input the optimal feature combination of the training set into the model obtained by the above training to get the brain age predicted by the diffusion index, and calculate the evaluation indicators of the predicted brain age and the actual age, including the mean absolute error MAE, the root mean square error RMSE, and the Pearson coefficient of the actual age and the predicted age, which are used to measure the accuracy and effectiveness of this brain age prediction model.

[0063] The content described in the embodiments of this specification is only an enumeration of the implementation forms of the inventive concept, and is only for illustrative purposes. The protection scope of the present invention should not be regarded as limited to the specific forms stated in this embodiment, and the protection scope of the present invention also extends to equivalent technical means that those of ordinary skill in the art can think of according to the inventive concept of the present invention.

Claims

1. A method for predicting brain age based on automatic recognition of fiber bundles, characterized in that: The method includes the following steps: Step 1: Perform image preprocessing on magnetic resonance diffusion tensor imaging, and the process is as follows: Use FSL tools to perform noise reduction, head motion correction, eddy current correction, distortion correction, and EPI correction operations on the image; Step 2: Trace the whole-brain fiber bundles, and the process is as follows: Use the UKF fiber tracking method with double tensors to trace the whole-brain fiber bundles to obtain a whole-brain fiber bundle map; Step 3: Automatically segment and identify the fiber bundles of the whole brain, and the process is as follows: Automatically segment the whole-brain fiber bundle map obtained in Step 2 using a cerebral white matter atlas. This method can segment small fiber bundles that are difficult to segment manually, and obtain 13 fiber bundles in each of the left and right brains and 6 commissural fibers; Step 4: Extract tensor values FA, MD, RD, and AD along the fiber bundles as white matter features, and the process is as follows: Use the BUAN algorithm to perform position-based equidistant sampling along the fiber bundle direction, sample the fiber bundle into a set number of segments, and calculate the average value of the diffusion tensor of each segment as the white matter feature; In the above Step 4, the process of automatically identifying the whole-brain fiber bundle map is as follows: First, divide the centroid of the atlas cluster in Step 3 into 100 segments along the length in the atlas space to obtain a template cluster. Then, for the cluster to be segmented, calculate the Euclidean distance between each point on each streamline of the cluster and the centroid of the template cluster, and assign the point to the nearest centroid segment; obtain the eigenvalues λ1, λ2, and λ3 in three directions in the three-dimensional space using the diffusion tensor matrix, and then calculate the fractional anisotropy FA, mean diffusivity MD, radial diffusivity RD, and axial diffusivity AD of each point; Step 5: Perform feature selection on the features extracted in Step 4, train an age prediction model, and use the test set to test and evaluate the generalization ability of the model, and analyze the fiber bundles that are more sensitive to age in the prediction model, and the process is as follows: Due to the problem of multicollinearity in the features extracted in Step 4, first perform recursive feature elimination on the features, and then train a brain age prediction model using ridge regression, and analyze the fiber bundles related to age changes according to the model weights.

2. The brain age prediction method based on automatic fiber bundle recognition according to claim 1, wherein In the above Step 3, the process of automatically identifying the whole-brain fiber bundle map is as follows: First, register the whole-brain fiber bundle map of each subject to the cerebral white matter atlas space, and cluster the whole-brain fiber bundles into 800 clusters according to the white matter partition generalization WMPG, inter-subject partition feasibility ISPV, and fiber bundle anatomical contour consistency TAPC. Each fiber in the whole-brain fiber bundle of each subject is assigned to the nearest atlas cluster in the atlas space; moreover, divide the fiber clusters of each subject into left hemisphere bundles, right hemisphere bundles, and commissural bundles, and further compare with the atlas bundles to remove abnormal fibers, and then identify the clusters with the same structure and function in the 800 clusters as fiber bundles with anatomical significance, such as the superior longitudinal fasciculus SLF.

3. The brain age prediction method based on automatic fiber bundle recognition according to claim 1 or 2, characterized in that, In the above Step 4, the fractional anisotropy FA, mean diffusivity MD, radial diffusivity RD, and axial diffusivity AD of each point are calculated as follows: AD = λ1 After dividing each cluster into 100 segments, calculate the average values of the diffusion tensor values FA, MD, RD, and AD of each segment as the white matter features of the brain age prediction model.

4. The brain age prediction method based on automatic fiber bundle recognition according to claim 1 or 2, characterized in that In the fifth step described above, the training and testing process of the brain age prediction model is as follows: 5.1) Feature Selection: Feature Selection In the first step, the Pearson correlation coefficient between each feature in the anatomically significant fiber bundle and the actual age is calculated first. A positive Pearson correlation coefficient indicates that the feature is positively correlated with the actual age, and a negative Pearson correlation coefficient indicates that the feature is negatively correlated with the actual age. The weights are calculated according to the correlation with age. The weight calculation formula is as follows: where n represents the total number of fibers contained in a bundle of fibers, and r i represents the Pearson coefficient between each white matter feature in the bundle and the actual age; Then, the weighted average of the diffusion tensor values of each bundle of fibers is obtained. There are 33 bundles of fibers and 4 diffusion tensor values in total, so there are 33 * 4 = 132 eigenvalue; since the problem of too high feature dimension will lead to model overfitting, the recursive feature elimination method is used to try various feature combinations, and the feature combination with the best model training effect is selected as the model input; 5.2) Feature normalization: After processing the feature outliers, the features are scaled so that the feature values are distributed between 0 and 1, and 80% of all samples are divided into the training set, and the remaining 20% is the test set; 5.3) Training the model: Using the feature combination obtained by feature selection as the eigenvalue and the actual age of the training set samples as the label, input into the ridge regression model for training. Five-fold cross-validation is adopted, and the evaluation indexes of the five validation sets are averaged. Grid search is used to find the model parameters that make the model effect the best; the coefficient of each item of this regression model is the weight of the feature. The larger the coefficient, the more sensitive the bundle of fibers is to age; 5.4) Testing the model: Inputting the optimal feature combination of the training set into the model obtained by the above training to get the brain age predicted by the diffusion index. Calculate the evaluation indexes between the predicted brain age and the actual age, including the mean absolute error MAE, the root mean square error RMSE, and the Pearson correlation coefficient between the actual age and the predicted age, which are used to measure the accuracy and effectiveness of this brain age prediction model.

Images