Brain age prediction method, device and equipment and storage medium

Abstract

The invention discloses a brain age prediction method, device and equipment and a storage medium, and relates to the technical field of image processing, and the method comprises the following steps: obtaining a T1w image and a DTI image of a brain; respectively preprocessing the T1w image and the DTI image to obtain a grey matter image and a 3D FA image; inputting the grey matter image and the 3D FA image into a brain age prediction model, and performing feature extraction on the grey matter image and the 3D FA image based on a feature extraction layer to obtain different deep features; fusing different deep features based on a feature fusion layer to obtain fused features; and mapping the fusion features based on the mapping layer to obtain a brain age prediction value. According to the method, the convolutional network is adopted to construct the brain age prediction model, only the T1w image and the DTI image need to be subjected to a simple preprocessing process, and compared with a traditional machine learning method, the process of complex preprocessing and descriptive feature extraction is omitted, and a large amount of time is saved.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a brain age prediction method, device, equipment and storage medium. Background Art

[0002] Normal aging can be thought of as a progression along a time trajectory. The biological age of an individual increases linearly over time, however, the predicted brain age obtained from neuroimaging endophenotypes using machine learning and deep learning techniques may be nonlinear because they depend on genetics, environment, and their interactions. The predicted brain age gap (predicted brain age - biological age) reflects the extent to which the brain deviates from the normal aging trajectory under a specific disease. An increase in the predicted brain age gap indicates an accelerated brain aging trajectory and an increased risk of neurodegeneration and other age-related problems, which has been widely revealed in neurological diseases. The study of brain age can also help in the diagnosis, treatment and prevention of related diseases in clinical practice. For example, through brain age prediction, the decline or abnormality of individual cognitive function can be detected early, which is helpful for early intervention and treatment, and delaying the process of cognitive decline. And it can help identify people with a higher risk of cognitive impairment and neurodegenerative diseases, which helps prevent and reduce the risk of disease. In addition, brain age prediction can be used as an indicator to evaluate the effectiveness of treatment, monitor the improvement of cognitive function during treatment, and help adjust treatment plans and improve treatment effects. Therefore, establishing a sufficiently accurate model to predict brain age has important neuroscience and clinical value.

[0003] Magnetic resonance imaging has the advantages of high resolution, rich information and no ionizing radiation damage, so it is widely used in medical diagnosis. In brain age prediction, T1w modality images are widely used because of their high spatial resolution and can carefully reflect the morphological characteristics of the anatomical structure in the brain (such as cortical thickness, volume of major brain regions, etc.). Diffusion tensor imaging (DTI) can provide information about white matter microstructure, such as the direction, connectivity and integrity of fiber bundles, which helps to study the organizational structure of brain nerve fibers. And by analyzing the changes in white matter fiber bundles, it helps to diagnose and screen neurodegenerative diseases such as Alzheimer's disease and Parkinson's disease at an early stage. At the same time, DTI can be used to study the changes in white matter microstructure during neural development and learning, which helps to understand brain development and learning mechanisms. In addition, these morphological characteristics have good clinical interpretability and can indicate the degree of neurodegeneration and the process of brain aging.

[0004] In recent years, many studies have obtained brain age with high predictive accuracy based on T1w and DTI images by using machine learning methods. Using these methods to determine brain aging patterns can help identify and quantify pathological changes in brain structure and function under different conditions (such as traumatic brain injury, schizophrenia, and Alzheimer's disease).

[0005] Current machine learning methods often require specialized and complex preprocessing procedures such as bias correction, segmentation, deformable registration, and coordination between scanners. These preprocessing steps are also not easily repeatable across different scanners, population cohorts, and MRI acquisition protocols. Machine learning methods also rely on descriptive features of brain regions obtained from complex and time-consuming image feature extraction, which results in excessively long prediction times for existing machine learning methods. Summary of the invention

[0006] The present invention provides a brain age prediction method, device, equipment and storage medium, which solves the problem that the current machine learning method requires a special and complex preprocessing process, and also relies on the descriptive features of the brain area obtained from complex and time-consuming feature extraction, so that the prediction time is too long when it is applied to brain age prediction.

[0007] In a first aspect, the present invention provides a method for predicting brain age, comprising the following steps:

[0008] Acquire T1w images and DTI images of the brain; T1w is T1 weighted, and DTI is diffusion tensor imaging;

[0009] T1w images and DTI images were preprocessed to obtain gray matter images and 3D FA images;

[0010] Constructing a brain age prediction model, including a feature extraction layer, a feature fusion layer and a mapping layer, wherein the feature extraction layer includes two identical network structures, each of which includes five feature extraction modules, wherein the first feature extraction module includes a 5x5x5 3D convolution layer and a 2x2x2 maximum pooling layer, the second feature extraction module, the third feature extraction module and the fourth feature extraction module each include a dense feature extraction block, a 1x1x1 3D convolution layer and an average pooling layer, and the fifth feature module includes a dense feature extraction block and an average pooling layer;

[0011] The gray matter image and the 3D FA image are input into the brain age prediction model. The feature extraction layer extracts features of the gray matter image and the 3DFA image respectively to obtain different deep features. The feature fusion layer fuses different deep features to obtain fused features. The mapping layer maps the fused features to obtain the brain age prediction value.

[0012] Preferably, the dense feature extraction block includes a 1x1x1 3D convolution layer, a 3x3x3 3D ​​convolution layer, two 3D batch normalization layers and two ReLU non-linear activation layers. The dense feature extraction blocks are closely interconnected in a feedforward manner within the block. The 3D convolution layer is used to extract features of different scales of the input image to obtain a feature image, and the 3D batch normalization layer is used to renormalize the feature image.

[0013] Preferably, the feature fusion layer includes a flattening operation and a splicing operation. The flattening operation flattens different deep features, and the splicing operation splices the flattened different deep features.

[0014] Preferably, the preprocessing of the T1w image and the DTI image to obtain the gray matter image and the 3D FA image respectively includes:

[0015] Convert the format of T1w images;

[0016] The skull removal tool is used to remove the skull of the T1w image after the format conversion to obtain a T1w skull-removed image;

[0017] The T1w image without skull is registered by linear registration tool to obtain T1w registered image;

[0018] Use the brain segmentation tool to segment the T1w registered image to obtain gray matter images and white matter images;

[0019] Perform eddy current correction on DTI images;

[0020] Extract b0 image from DTI image, perform brain peeling on b0 image to obtain mask image;

[0021] Calculate the anisotropy scores of the mask image to obtain the FA image;

[0022] The FA image is processed by TBSS to obtain a 3D FA image.

[0023] Preferably, the weights of the brain age prediction model are trained by a loss function, and the weights are optimized by an Adam optimizer; the loss function is the mean square error between the predicted brain age and the actual age.

[0024] In a second aspect, the present invention further provides a brain age prediction device, characterized in that it comprises:

[0025] An acquisition module, used to acquire T1w images and DTI images of the brain; the T1w is T1 weighted, and the DTI is diffusion tensor imaging;

[0026] A preprocessing module is used to preprocess the T1w image and the DTI image to obtain a gray matter image and a 3D FA image;

[0027] A construction module, for constructing a brain age prediction model, comprising a feature extraction layer, a feature fusion layer and a mapping layer, wherein the feature extraction layer comprises two identical network structures, each of which comprises five feature extraction modules, wherein the first feature extraction module comprises a 5x5x5 3D convolution layer and a 2x2x2 maximum pooling layer, the second feature extraction module, the third feature extraction module and the fourth feature extraction module each comprise a dense feature extraction block, a 1x1x1 3D convolution layer and an average pooling layer, and the fifth feature extraction module comprises a dense feature extraction block and an average pooling layer;

[0028] The prediction module is used to input the gray matter image and the 3D FA image into the brain age prediction model, extract features of the gray matter image and the 3D FA image based on the feature extraction layer to obtain different deep features; fuse different deep features based on the feature fusion layer to obtain fused features; and map the fused features based on the mapping layer to obtain the brain age prediction value.

[0029] In a third aspect, the present invention further provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned brain age prediction method when executing the program.

[0030] In a fourth aspect, the present invention further provides a computer-readable storage medium, wherein the storage medium stores a computer program, and the computer program implements the above-mentioned brain age prediction method when executed by a processor.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] The present invention first obtains T1w images and DTI images of the brain, and constructs a brain age prediction model based on a convolutional network. Based on the constructed brain age prediction model, the T1w image and DTI image are preprocessed to obtain a gray matter image and a 3D FA image. The present invention uses a convolutional network to construct a brain age prediction model, and only requires a simple preprocessing process for the T1w image and the DTI image. Compared with the traditional machine learning method, the complex preprocessing and descriptive feature extraction process are omitted, saving a lot of time. In addition, the brain age prediction model combines deep extraction features of different modalities, improves the fitting generalization ability of the ensemble model, and thus improves the accuracy of brain age prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0034] Figure 1 is a flow chart of preprocessing of T1w images of the present invention;

[0035] Figure 2 is a flow chart of preprocessing of DTI images of the present invention;

[0036] Figure 3 is a schematic diagram of the structure of the brain age prediction model of the present invention;

[0037] Figure 4 It is a schematic diagram of the structure of the dense feature extraction block of the present invention; DETAILED DESCRIPTION

[0038] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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 creative work are within the scope of protection of the present invention.

[0039] The present invention provides a method for predicting brain age, which specifically predicts brain age by using a convolutional neural network, and comprises the following steps:

[0040] Step 1: Obtain T1w images and DTI images of the brain, T1w is T1 weighted, and preprocess the T1w images and DTI images.

[0041] First, T1w images and DTI images of the brain are collected. Predicting brain age through convolutional neural networks only requires a relatively simple preprocessing process for the T1w images and DTI images. It does not require a special and complex preprocessing process like existing machine learning methods, which facilitates the clinical use of deep learning methods.

[0042] Figure 1The figure shows the preprocessing operation process of T1w image. Because there is a high correlation between the characteristics of gray matter image and the brain aging process, the preprocessing of T1w image is the extraction of gray matter image. It includes steps such as image format conversion, image skull removal, image registration, and image gray matter and white matter segmentation. All preprocessing steps of the present invention are carried out in FMRIB's Software Library (FSL) (FMRIB software library). First, the T1w image is converted into a format suitable for subsequent processing, and the T1w image after the converted format is skull removed using FSL BET (skull removal tool), and the brain tissue part is retained to obtain the T1w skull removal image. Afterwards, the T1w skull removal image is registered to the MNI standard space using FSL FLIRT (linear registration tool) to ensure that all images are in the same coordinate system to obtain the T1w registered image. Finally, the T1w registered image is segmented into gray matter image and white matter image by FSL FAST tool (FSL brain segmentation tool).

[0043] Figure 2 The figure shows a series of preprocessing operations on the DTI raw data, including extraction of b0 image, brain stripping, eddy current correction, tensor index calculation, nonlinear registration, and creation of average FA image and skeleton. The b0 image is extracted from the DTI raw data, and then the original data is brain stripped, that is, the image of non-brain tissue structure and skull is stripped to obtain the mask image after brain stripping. Distortion caused by eddy current and head movement is the most common artifact, so eddy current correction is required. Then tensor indexes such as fractional anisotropy (FA) and mean diffusivity (MD) are calculated. The obtained FA image is processed by TBSS, which includes three steps: data preparation, nonlinear registration, and creation of average FA image and skeleton, to obtain a 3D FA image of each subject in the standard MNI152 space. The gray matter image and 3D FA image obtained by preprocessing will be input into the deep fusion model for brain age prediction.

[0044] Step 2: Input the gray matter image and 3D FA image into the brain age prediction model to obtain the brain age prediction value.

[0045] Figure 3Shown in the figure is the structural diagram of the deep fusion model based on multimodal features. The deep fusion model uses a gray matter image with a volume of 182x218x182 and a 3D FA image after data preprocessing as input. The proposed deep fusion model based on multimodal features consists of two identical modified 3D-Densenet network models. The modified 3D-Densenet network model consists of five feature extraction modules. The first feature extraction module consists of a 5x5x5 3D convolution layer (conv) and a 2x2x2 maximum pooling layer (Max Pooling). The second, third and fourth feature extraction modules consist of a dense feature extraction block (Dense block), a 1x1x1 3D convolution layer and an average pooling layer (AveragePool). The fifth feature extraction module consists of a dense feature extraction block and an average pooling layer. Refer to Figure 4 The dense feature extraction block contains a 1x1x1 3D convolution layer, a 3x3x3 3D ​​convolution layer, two 3D batch normalization layers (BatchNorm3D) and two ReLU nonlinear activation layers. The dense feature extraction blocks are closely interconnected in a feedforward manner within the block. Among them, the number of layers of the 1st to 4th dense feature extraction blocks are 3, 6, 12 and 8 respectively, and the growth rate of the dense feature extraction block is 48. After the deep extraction features based on gray matter and the deep extraction features based on 3D FA white matter are extracted by the 3D-Densenet network model, the deep extraction features of different modalities are fused using the Flatten and Concatenate methods, and finally the fused feature map is used through a fully connected layer to obtain the final brain age prediction value.

[0046] Among them, the convolution layer is used to extract the features of the input image, and the batch normalization layer is used to renormalize the feature image to avoid the gradient disappearance caused by the extremely deep network, so that the network can be well trained. The maximum pooling layer and the average pooling layer can effectively reduce the size of the feature image, thereby speeding up the calculation speed and preventing overfitting. The fully connected layer is responsible for converting the output three-dimensional feature map into a one-dimensional vector, thereby realizing the end-to-end (from three-dimensional image to brain age prediction value) learning process, while reducing the influence of feature position on the result, and improving the robustness of the entire network. ReLU nonlinear activation helps to enhance the nonlinear ability of the model and accelerate model training. The mean square error (MSE) between the brain age prediction value and the actual age is used as the loss function to train the weights of the deep learning model network. The Adam optimizer is used to optimize the neural network model. The Adam algorithm is a first-order optimization algorithm that can iteratively update the neural network weights based on the training data, making it easy to implement, efficient in calculation, and requires less memory, and can ensure the accuracy of the prediction results. The He initialization strategy is used to initialize the weights of the deep learning network. Each trainable node in the network uses L2 weight decay (λ=5e -5 ) for regularization. An early stopping strategy for model checkpointing is also used to obtain the best model during training. In addition, data augmentation generates new training images by applying coordinate transformations (including 3D rotation and 3D translation) to random subsets of the training data, thereby increasing the number of datasets, reducing the risk of overfitting, and improving the robustness of the model.

[0047] Compared with traditional machine learning methods, the brain age prediction model proposed in the present invention omits the complex feature extraction process and only requires a relatively simple preprocessing process for the image, saving a lot of time. In addition, by using deep learning technology combined with deep extraction features of different modalities, it can overcome the problems of data sparsity, parameter priori and model bias in traditional brain age models, further improving the fitting and generalization ability of the ensemble model, thereby improving the accuracy and robustness of brain age prediction.

[0048] Based on the same concept, the present invention also provides a brain age prediction device, including an acquisition module, a preprocessing module, a construction module and a prediction module.

[0049] The acquisition module is used to acquire T1w images and DTI images of the brain; T1w is T1 weighted, and DTI is diffusion tensor imaging.

[0050] The preprocessing module is used to preprocess the T1w image and the DTI image respectively to obtain the gray matter image and the 3D FA image.

[0051] The building module is used to construct a brain age prediction model, including a feature extraction layer, a feature fusion layer and a mapping layer. The feature extraction layer includes two identical network structures, each of which includes five feature extraction modules. The first feature extraction module includes a 5x5x5 3D convolution layer and a 2x2x2 maximum pooling layer. The second feature extraction module, the third feature extraction module and the fourth feature extraction module all include a dense feature extraction block, a 1x1x1 3D convolution layer and an average pooling layer. The fifth feature module includes a dense feature extraction block and an average pooling layer.

[0052] The prediction module is used to input the gray matter image and the 3D FA image into the brain age prediction model, perform feature extraction on the gray matter image and the 3D FA image based on the feature extraction layer to obtain different deep features; fuse different deep features based on the feature fusion layer to obtain fused features; and map the fused features based on the mapping layer to obtain the brain age prediction value.

[0053] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the above-mentioned brain age prediction method is implemented when the processor executes the program.

[0054] The present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the above-mentioned brain age prediction method is implemented.

[0055] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0056] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for predicting brain age, characterized in that: The following steps are involved: Acquire T1w images and DTI images of the brain; T1w is T1 weighted, and DTI is diffusion tensor imaging; T1w images and DTI images were preprocessed to obtain gray matter images and 3D FA images; Constructing a brain age prediction model, including a feature extraction layer, a feature fusion layer and a mapping layer, wherein the feature extraction layer includes two identical network structures, each of which includes five feature extraction modules, wherein the first feature extraction module includes a 5x5x5 3D convolution layer and a 2x2x2 maximum pooling layer, the second feature extraction module, the third feature extraction module and the fourth feature extraction module each include a dense feature extraction block, a 1x1x1 3D convolution layer and an average pooling layer, and the fifth feature module includes a dense feature extraction block and an average pooling layer; The gray matter image and the 3D FA image are input into the brain age prediction model, and the gray matter image and the 3D FA image are respectively subjected to feature extraction based on the feature extraction layer to obtain different deep features; Based on the feature fusion layer, different deep features are fused to obtain fused features; The fused features are mapped based on the mapping layer to obtain the predicted value of brain age.

2. A method for predicting brain age according to claim 1, characterized in that: The dense feature extraction block includes a 1x1x1 3D convolution layer, a 3x3x3 3D ​​convolution layer, two 3D batch normalization layers and two ReLU non-linear activation layers. The dense feature extraction blocks are closely interconnected in a feedforward manner within the block. The 3D convolution layer is used to extract features of different scales of the input image to obtain a feature image. The 3D batch normalization layer is used to renormalize the feature image.

3. A method for predicting brain age according to claim 1, characterized in that: The feature fusion layer includes a flattening operation and a splicing operation. The flattening operation flattens different deep features, and the splicing operation splices the flattened different deep features.

4. A method for predicting brain age as claimed in claim 1, characterized in that: The T1w image and the DTI image are preprocessed respectively to obtain a gray matter image and a 3D FA image, including: Convert the format of T1w images; The skull removal tool is used to remove the skull of the T1w image after the format conversion to obtain a T1w skull-removed image; The T1w image without skull is registered by linear registration tool to obtain T1w registered image; Use the brain segmentation tool to segment the T1w registered image to obtain gray matter images and white matter images; Perform eddy current correction on DTI images; Extract b0 image from DTI image, perform brain peeling on b0 image to obtain mask image; Calculate the anisotropy scores of the mask image to obtain the FA image; The FA image is processed by TBSS to obtain a 3D FA image.

5. A method for predicting brain age according to claim 1, characterized in that: The weights of the brain age prediction model are trained by a loss function, and the weights are optimized by an Adam optimizer; the loss function is the mean square error between the predicted brain age and the actual age.

6. A brain age prediction device, characterized in that: include: An acquisition module, used to acquire T1w images and DTI images of the brain; The T1w is T1 weighted, and DTI is diffusion tensor imaging; A preprocessing module is used to preprocess the T1w image and the DTI image to obtain a gray matter image and a 3D FA image; A construction module, for constructing a brain age prediction model, comprising a feature extraction layer, a feature fusion layer and a mapping layer, wherein the feature extraction layer comprises two identical network structures, each of which comprises five feature extraction modules, wherein the first feature extraction module comprises a 5x5x5 3D convolution layer and a 2x2x2 maximum pooling layer, the second feature extraction module, the third feature extraction module and the fourth feature extraction module each comprise a dense feature extraction block, a 1x1x1 3D convolution layer and an average pooling layer, and the fifth feature extraction module comprises a dense feature extraction block and an average pooling layer; A prediction module is used to input the gray matter image and the 3D FA image into the brain age prediction model, and extract features from the gray matter image and the 3D FA image based on the feature extraction layer to obtain different deep features; Based on the feature fusion layer, different deep features are fused to obtain fused features; The fused features are mapped based on the mapping layer to obtain the predicted value of brain age.

7. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method for predicting brain age described in any one of claims 1 to 5 is implemented.

8. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the brain age prediction method according to any one of claims 1 to 5 is implemented.

Images