Cascade polarization self-attention-based brain age prediction method
By introducing a cascaded polarized self-attention feature enhancement module into the brain age prediction network, the problem that existing methods are difficult to capture global information and channel dependencies is solved, and more accurate brain age prediction is achieved and prediction errors are reduced.
Patent Information
- Application Number
- CN202510171258.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-17
- Publication Date
- 2025-06-06
AI Technical Summary
Existing brain age prediction methods are difficult to effectively capture global information and channel dependencies, and are insufficient sensitivity to individual pixels, resulting in high brain age prediction errors.
A brain age prediction method based on cascaded polarized self-attention is adopted. A network model consisting of 5 feature extraction modules, 1 nonlinear module and 1 regression module is constructed, and a polarized self-attention feature enhancement module is introduced into each module to form a cascade structure to enhance feature extraction capabilities.
It significantly reduces brain age prediction errors, improves the model's ability to capture contextual characteristics, and can more accurately predict individual brain age, thereby assisting doctors in early intervention in neurodegenerative diseases.
Smart Images
Figure CN120107198A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of medical image processing, and specifically is a brain age prediction method based on cascaded polarization self-attention. Background Art
[0002] The brain is one of the most complex and vital organs in the human body, and its aging process is closely related to cognitive function, emotional state, and behavioral performance. Brain age refers to the predicted age of an individual derived from neuroimaging data in a machine learning framework, which may represent a biomarker of the potential age of the brain. The gap between the biological age of the brain and the chronological age is called the "brain age gap", which is a valuable biomarker of brain health. A positive brain age gap is associated with decreased cognitive ability and an increased risk of neurodegenerative diseases, such as Alzheimer's disease and mild cognitive impairment. Therefore, accurate prediction of brain biological age is of great significance for early detection and intervention in the progression of neurodegenerative diseases.
[0003] As a non-invasive technology, magnetic resonance imaging (MRI) provides high contrast and high spatial resolution in image acquisition, making it a widely used imaging method in brain age prediction. With the continuous development of machine learning and deep learning technologies, researchers have built a brain age prediction model by combining MRI image data with machine learning or deep learning technologies. Traditional machine learning methods rely on manual experience and expertise to manually extract features from image data, and then use these features to build prediction models. Although the requirements for the number of samples are low, their performance may be limited when faced with complex tasks. Deep learning can automatically learn feature representations from brain image data, reducing the difficulty of feature extraction, but its demand for large amounts of data and long training cycles still require improvement in its stability and generalization ability. Combining machine learning, deep learning and other algorithms, researchers can extract useful features from changes in brain structure and function to infer an individual's brain age. For healthy individuals, brain age is usually close to actual age; if the brain age is high, it indicates that the brain is aging faster and there may be certain health risks. Therefore, brain age, as a quantitative indicator, can not only reflect the developmental state of the brain, but also reveal the potential relationship between brain aging and neurological diseases.
[0004] As an important indicator for the prevention and health management of neurodegenerative diseases, brain age gap is gaining increasing attention. Therefore, it is of great significance to propose a model that can accurately predict the brain age of healthy people. By comparing the brain age of an individual with his or her actual age, researchers can identify those at high risk of abnormally accelerated brain aging and develop targeted preventive measures and treatment plans for them. Summary of the invention
[0005] The present invention proposes a brain age prediction method based on cascaded polarized self-attention, which aims to effectively solve the problems that existing methods are difficult to capture global information and channel dependencies, and have insufficient sensitivity to single pixels, reduce brain age prediction errors, and thus assist doctors in intervening in the treatment of neurodegenerative diseases in advance.
[0006] To achieve the above object, the present invention provides the following technical solution: a brain age prediction method based on cascaded polarization self-attention, comprising the following steps:
[0007] Step S1: preprocessing the original brain T1-weighted magnetic resonance image to obtain a preprocessed brain image dataset;
[0008] This included nonlinear registration of images to standard MNI space, brain extraction, and voxel brightness normalization (i.e., subtracting the mean voxel value of the brain region and dividing by the standard deviation).
[0009] Step S2: construct a brain age prediction network model based on cascaded polarized self-attention;
[0010] In step S2, the proposed network consists of 5 feature extraction modules, 1 nonlinear module and 1 regression module. In the first six modules, a polarized self-attention feature enhancement module is introduced at the end of each module to form a cascade structure, which effectively strengthens the features of each layer.
[0011] Each feature extraction module consists of a convolutional layer, a batch normalization layer, a maximum pooling layer, and a ReLU activation layer.
[0012] The nonlinear module consists of a 3D convolutional layer, a batch normalization layer, and a ReLU activation layer.
[0013] The regression module consists of an average pooling layer, a Dropout layer, and a 1×1×1 convolutional layer.
[0014] The polarization self-attention feature enhancement module simulates the filtering principle of optical lenses through polarization filtering operations, selectively highlights and suppresses features, and significantly enhances the model's ability to capture contextual features. The polarization self-attention feature enhancement module uses two sub-modules, the polarization channel self-attention module and the polarization space self-attention module, to enhance information in channels and spaces respectively to capture global and local dependencies.
[0015] The polarized channel self-attention module can capture global context features and effectively calculate attention scores to enhance channel features. First, the module calculates the Query and Value vectors of each attention point, as well as the global channel Key vector. Then, the matrix Q formed by the Query vectors of all attention points isch The matrix K formed with the global channel Key vector ch Multiply them together to get the channel attention score matrix Score ch Finally, the matrix Score ch and the matrix V formed by the Value vector ch Multiply them together to get the feature map after enhancing the channel information. The input of the polarization channel self-attention module is the output feature map of the feature extraction module. Where C is the number of channels of the input image, and X, Y, and Z are the height, width, and depth of the image, respectively. The channel attention score matrix is defined as:
[0016] Q ch =R q (M q (I)),
[0017] K ch =F soft (R k (M k (I))),
[0018] Score ch =F sig (F ln (M z (Q ch ×K ch ))),
[0019] Among them, M q , M k , M z They are depth-wise separable convolution, R q , R k Represents different Reshape operations, F soft is the softmax operation, F ln Representation layer normalization operation, F sig is the sigmoid activation function. The output of the polarized channel self-attention module is the channel enhanced feature map
[0020] V ch =M v (I),
[0021] I ch =V ch ⊙Score ch ,
[0022] Among them, M v is the convolutional layer, and ⊙ is the element-wise product.
[0023] The polarized spatial self-attention module first calculates the matrix Q corresponding to the Query vector, Value vector and global space Key vector of each focus point. sp , Q sp , K sp Then, the matrix K sp With the matrix Q sp Multiply them together to get the spatial attention score matrix Score sp Finally, the matrix Score sp With the matrix V sp Multiply element by element to get the spatially enhanced feature map. The input of the polarized spatial self-attention module is the channel-enhanced feature map Spatial attention score matrix Score sp Defined as:
[0024] Q sp =R q (M q (I ch )),
[0025] K sp =F soft (R k (F GA (M k (I ch )))),
[0026] Score sp =F sig ( R z(K sp ×Q sp )),
[0027] Here M q 、M k , R q , R k , R z 、F soft and F sig As described above, the same functionality is maintained in the network architecture. GA Represents the global average operation. The output of the polarized spatial self-attention module is the spatially enhanced feature map
[0028] V sp =M v (I ch ),
[0029] I sp =V sp ⊙Score sp ,
[0030] Among them, Mv is a convolutional layer, and ⊙ is an element-wise product.
[0031] Step S3: Input the preprocessed brain magnetic resonance image into the proposed brain age prediction network for training to obtain a trained brain age prediction network model.
[0032] Step S4: Use the trained network model to predict the brain image data of the test set to obtain the predicted brain age of the subject. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 It is an implementation flow chart of the present invention.
[0034] Figure 2 This is a brain age prediction network structure diagram of the present invention.
[0035] Figure 3 This is a structural diagram of the polarization channel self-attention module of the present invention.
[0036] Figure 4 This is a structural diagram of the polarized spatial self-attention module of the present invention. DETAILED DESCRIPTION
[0037] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.
[0038] Embodiment: A method for predicting brain age based on cascaded polarization self-attention comprises the following steps:
[0039] refer to Figure 1 The steps of implementing the invention include: (1) preprocessing the original brain T1-weighted magnetic resonance images to obtain a preprocessed brain image data set; (2) constructing a brain age prediction network model based on cascaded polarization self-attention; (3) inputting the preprocessed brain magnetic resonance images into the proposed brain age prediction network for training to obtain a trained brain age prediction network model; (4) using the trained network model to predict the brain image data of the test set to obtain the predicted brain age of the subject.
[0040] refer to Figure 2 , is the structural diagram of the proposed brain age prediction network. The proposed network consists of 5 feature extraction modules, 1 nonlinear module and 1 regression module. In the first six modules, a polarized self-attention feature enhancement module is introduced at the end of each module to form a cascade structure, which effectively strengthens the features of each layer.
[0041] Each feature extraction module consists of a convolutional layer, a batch normalization layer, a maximum pooling layer, and a ReLU activation layer.
[0042] The nonlinear module consists of a 3D convolutional layer, a batch normalization layer, and a ReLU activation layer.
[0043] The regression module consists of an average pooling layer, a Dropout layer, and a 1×1×1 convolutional layer.
[0044] The polarization self-attention feature enhancement module simulates the filtering principle of optical lenses through polarization filtering operations, selectively highlights and suppresses features, and significantly enhances the model's ability to capture contextual features. The polarization self-attention feature enhancement module uses two sub-modules, the polarization channel self-attention module and the polarization space self-attention module, to enhance information in channels and spaces respectively to capture global and local dependencies.
[0045] refer to Figure 3 , is the structure diagram of the polarized channel self-attention module, which can capture global context features and effectively calculate attention scores to enhance channel features. First, the module calculates the Query and Value vectors of each focus point, as well as the global channel Key vector. Then, the matrix Q formed by the Query vectors of all focus points is ch The matrix K formed with the global channel Key vector ch Multiply them together to get the channel attention score matrix Score ch Finally, the matrix Score ch and the matrix V formed by the Value vector ch Multiply them together to get the feature map after enhancing the channel information. The input of the polarization channel self-attention module is the output feature map of the feature extraction module. Where C is the number of channels of the input image, and X, Y, and Z are the height, width, and depth of the image, respectively. The channel attention score matrix is defined as:
[0046] Q ch =R q (M q (I)),
[0047] K ch =F soft (R k (M k (I))),
[0048] Score ch =F sig (F ln (M z (Q ch ×K ch ))),
[0049] Among them, M q , M k, M z They are depth-wise separable convolution, R q , R k Represents different Reshape operations, F soft is the softmax operation, F ln Representation layer normalization operation, F sig is the sigmoid activation function. The output of the polarized channel self-attention module is the channel enhanced feature map
[0050] V ch =M v (I),
[0051] I ch =V ch ⊙Score ch ,
[0052] Among them, M v is the convolutional layer, and ⊙ is the element-wise product.
[0053] refer to Figure 4 , is the structure diagram of the polarized space self-attention module. This module first calculates the Query vector, Value vector and the matrix Q corresponding to the global space Key vector of each focus point. sp , V sp , K sp Then, the matrix K sp With the matrix Q sp Multiply them together to get the spatial attention score matrix Score sp Finally, the matrix Score sp With the matrix V sp Multiply element by element to get the spatially enhanced feature map. The input of the polarized spatial self-attention module is the channel-enhanced feature map Spatial attention score matrix Score sp Defined as:
[0054] Q sp =R q (M q (I ch )),
[0055] K sp =F soft (R k (F GA (M K k(I ch )))),
[0056] Score sp =F sig (R z (Ksp ×Q sp )),
[0057] Here M q 、M k , R q , R k , R z 、F soft and F sig As described above, the same functionality is maintained in the network architecture. GA Represents the global average operation. The output of the polarized spatial self-attention module is the spatially enhanced feature map
[0058] V sp =M v (I ch ),
[0059] I sp =V sp ⊙Score sp ,
[0060] Among them, M v is a convolutional layer, and ⊙ is an element-wise product.
[0061] The above contents are further detailed descriptions of the present invention in combination with the best embodiment, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. It should be understood by those skilled in the art that various modifications can be made in the details without departing from the scope of the appended claims, and all should be deemed to belong to the protection scope of the present invention.
Claims
1. A brain age prediction method based on cascaded polarization self-attention, characterized in that: The following steps are included: Step S1: preprocessing the original brain T1-weighted magnetic resonance image to obtain a preprocessed brain image dataset; Step S2: construct a brain age prediction network model based on cascaded polarized self-attention; Step S3: inputting the preprocessed brain magnetic resonance image into the proposed brain age prediction network for training to obtain a trained brain age prediction network model; Step S4: Use the trained network model to predict the brain image data of the test set to obtain the predicted brain age of the subject.
2. The method for predicting brain age based on cascaded polarized self-attention according to claim 1, characterized in that: In step S1 , the preprocessing steps include non-linear registration of the image to the standard MNI space, brain extraction, and normalization of voxel intensity values (ie, subtracting the mean voxel value of the brain region and dividing by the standard deviation).
3. The brain age prediction method based on cascaded polarized self-attention according to claim 1, characterized in that: In step S2, the proposed network consists of 5 feature extraction modules, 1 nonlinear module and 1 regression module. In the first six modules, a polarized self-attention feature enhancement module is introduced at the end of each module to form a cascade structure, which effectively strengthens the features of each layer.
4. The method for predicting brain age based on cascaded polarized self-attention according to claim 3, characterized in that: Each feature extraction module consists of a convolutional layer, a batch normalization layer, a maximum pooling layer, and a ReLU activation layer.
5. The method for predicting brain age based on cascaded polarized self-attention according to claim 3, characterized in that: The nonlinear module consists of a 3D convolutional layer, a batch normalization layer, and a ReLU activation layer.
6. The method for predicting brain age based on cascaded polarized self-attention according to claim 3, characterized in that: The regression module consists of an average pooling layer, a Dropout layer, and a 1×1×1 convolutional layer.
7. The method for predicting brain age based on cascaded polarized self-attention according to claim 3, characterized in that: Polarized self-attention feature enhancement module: This module simulates the filtering principle of optical lenses through polarized filtering operations, selectively highlights and suppresses features, and significantly enhances the model's ability to capture contextual features. The polarized self-attention feature enhancement module uses two sub-modules, the polarized channel self-attention module and the polarized spatial self-attention module, to enhance information in channels and spaces respectively to capture global and local dependencies.
8. The method for predicting brain age based on cascaded polarized self-attention according to claim 7, characterized in that: The polarized channel self-attention module can capture global context features and effectively calculate attention scores to enhance channel features. First, the module calculates the Query and Value vectors of each attention point, as well as the global channel Key vector. Then, the matrix Q formed by the Query vectors of all attention points is ch The matrix K formed with the global channel Key vector ch Multiply them together to get the channel attention score matrix Score ch Finally, the matrix Score ch and the matrix V formed by the Value vector ch Multiply them together to get the feature map after enhancing the channel information. The input of the polarization channel self-attention module is the output feature map of the feature extraction module. Where C is the number of channels of the input image, and X, Y, and Z are the height, width, and depth of the image, respectively. The channel attention score matrix is defined as: Q ch =R q (M q (I)), K ch =F soft (R k (M k (I))), Score ch =F sig (F ln (M z (Q ch ×K ch ))), Among them, M q , M k , M z They are depth-wise separable convolution, R q , R k Represents different Reshape operations, F soft is the softmax operation, F ln Representation layer normalization operation, F sig is the sigmoid activation function. The output of the polarized channel self-attention module is the channel enhanced feature map V ch =M v (I), I ch =V ch ⊙Score ch , Among them, M v is the convolutional layer, and ⊙ is the element-wise product.
9. The method for predicting brain age based on cascaded polarized self-attention according to claim 7, characterized in that: Polarized spatial self-attention module. First, this module calculates the matrix Q corresponding to the Query vector, Value vector and global space Key vector of each focus point. sp , V sp , K sp Then, the matrix K sp With the matrix Q sp Multiply them together to get the spatial attention score matrix Score sp Finally, the matrix Score sp With the matrix V sp Multiply element by element to get the spatially enhanced feature map. The input of the polarized spatial self-attention module is the channel-enhanced feature map Spatial attention score matrix Score sp Defined as: Q sp =R q (M q (I ch )), K sp =F soft (R k (F GA (M k (I ch )))), Score sp =F sig (R z (K sp ×Q sp )), Here M q 、M k , R q , R k , R z 、F soft and F sig As described above, the same functionality is maintained in the network architecture. GA Represents the global average operation. The output of the polarized spatial self-attention module is the spatially enhanced feature map V sp =M v (I ch ), I sp =V sp ⊙Score sp , Among them, M v is the convolutional layer, and ⊙ is the element-wise product.