A method, system, device and medium for predicting brain age

Abstract

This invention discloses a brain age prediction method, system, device, and medium, relating to the field of biomedical image analysis technology. First, it accurately captures precise local detail features representing the cerebral cortex and brain tissue. Then, based on a windowed multi-head self-attention mechanism, it obtains the dependencies between different local regions of the image. It also introduces a channel attention mechanism, learning the importance weights of different feature channels and applying these weights to feature fusion to accurately capture long-distance dependencies and subtle structural features. Next, based on a cross-attention mechanism, it guides global detail features to focus on regions rich in local features, and guides local detail features to focus on the most relevant local regions, thereby uncovering deeper levels of detail features. Finally, it dynamically adjusts the contribution ratio of local detail features and global contextual detail features in the final decision and fuses them into the final feature, enabling a deeper perception of both local and global detail features.

Description

Technical Field

[0001] This invention relates to the field of biomedical image analysis technology, and in particular to a brain age prediction method, system, device and medium. Background Technology

[0002] "Brain age" is a biomarker that has attracted much attention in the fields of neuroscience and clinical medicine in recent years. It uses machine learning models built from neuroimaging data (especially structural magnetic resonance imaging, sMRI) to predict the physiological age of an individual's brain. The difference between the predicted brain age and the individual's actual age is called the "brain age gap" (BAG). Numerous studies have shown that a positive brain age gap (i.e., the predicted brain age is greater than the actual age) is closely related to the accelerated aging process of the brain and is an effective predictor of various neurological and mental diseases, such as Alzheimer's disease, Parkinson's disease, schizophrenia, and depression. Therefore, developing a high-precision brain age prediction technology is not only of great scientific value for understanding the health trajectory of an individual's brain and realizing early risk warning of neurodegenerative diseases, but also provides a new objective standard for the quantitative evaluation of clinical intervention effects, and has broad application prospects.

[0003] Currently, brain age prediction methods based on deep learning are commonly used, mainly including methods based on three-dimensional convolutional neural networks (3D-CNN) and methods based on visual Transformers (ViT). The 3D-CNN method can effectively extract subtle structural features (such as gray-white matter boundaries, cortical folds, etc.) in sMRI through local sliding of the convolution kernel. However, due to the limitation of the local receptive field of the convolution operation, it is difficult to effectively capture long-distance dependencies between different regions of the brain, and its ability to perceive global details is weak when perceiving local details. The ViT method can capture global dependencies in images through self-attention mechanism, but its method of processing 3D images into blocks destroys the spatial continuity at the voxel level, and its ability to perceive local details is weak when modeling global information.

[0004] Therefore, current deep learning-based brain age prediction methods are unable to simultaneously capture the precise long-distance dependencies and subtle structural features between different brain regions, making it difficult to simultaneously perceive local and global details at a deeper level. Summary of the Invention

[0005] This invention provides a brain age prediction method, system, device, and medium that can solve the problems existing in the prior art.

[0006] This invention provides a brain age prediction method, comprising the following steps:

[0007] Acquire sMRI data of the target subject;

[0008] Extract local detailed features from sMRI data that characterize the boundary texture of the cerebral cortex and the morphological features of brain tissue;

[0009] The sMRI data of nuclear magnetic resonance imaging (MRI) images were sliced ​​into image sequences, and features were extracted from each image sequence under different feature channels. During feature extraction, the dependencies between local regions of different images were obtained based on a windowed multi-head self-attention mechanism, and the importance weights of different feature channels were generated during global spatial feature extraction based on a channel self-attention mechanism. Based on the dependencies and the importance weights of different feature channels, the features extracted from different feature channels were weighted and fused to obtain global detailed features.

[0010] Based on the cross-attention mechanism, global detail features are used as keys for attention calculation, global attention is obtained and weighted fusion is performed to obtain global detail features rich in local features, and local detail features are used as queries for attention calculation, local attention is obtained and weighted fusion is performed to obtain local detail features enhanced in local features.

[0011] The original global detail features and local detail features are concatenated to obtain concatenated features. The concatenated features are then passed through a fully connected layer and a sigmoid activation function to obtain dynamic gating weights. Based on the dynamic gating weights, the global detail features rich in local features and the local detail features enhanced by local features are dynamically weighted and fused to obtain fused features. The brain age of the target measurer is predicted based on the fused features.

[0012] Preferably, the fusion feature is obtained through a brain age prediction model, which includes a cascaded lightweight 3D-CNN module, a Swing Transformer module, and a cross-attention gating fusion module.

[0013] The lightweight 3D-CNN module consists of 6 basic units connected in series. Each basic unit contains a 3D convolutional layer with a kernel size of 3x3x3 and a max pooling layer with a kernel size of 2x2x2, which is used to extract local detail features.

[0014] The Swin Transformer module is based on the Swin Transformer architecture and introduces a channel self-attention mechanism in parallel in each windowed multi-head self-attention block to extract global detail features.

[0015] The cross-attention gating fusion module includes cascaded linear layers, a cross-attention mechanism, and a fully connected layer. The activation function is the Sigmoid function, which is used to extract global detail features rich in local features and local detail features enhanced by local features, and to dynamically weight and fuse the global detail features rich in local features and the local detail features enhanced by local features.

[0016] Preferably, the acquisition of the local detail features includes:

[0017] Feature extraction is performed using a 3x3x3 convolutional layer to obtain a 3D feature map. This 3D feature map is then passed through a 2x2x2 max pooling layer to aggregate it into a 1D deep convolutional feature vector. ;

[0018] Among them, for the third-dimensional convolutional layer The 3D convolution operation of the layer outputs a feature map. any voxel on The value is represented as:

[0019] ;

[0020] in: Indicates the input feature map; and This indicates the kernel weights and biases of this layer; This represents the activation function.

[0021] Preferably, the acquisition of the global detailed features includes:

[0022] For an input image sequence ,in For sequence length, As the feature dimension, self-attention computation is represented as:

[0023] ;

[0024] in: These represent the query, key, and value matrices obtained through linear transformation, respectively.

[0025] Windowed multi-head self-attention restricts it to non-overlapping local windows and enables cross-window information interaction through window shifting SW-MSA;

[0026] Based on information interaction, a channel attention weight is introduced. Element-wise multiplication with the output features is performed to enhance the feature representation, resulting in a one-dimensional deep long-range feature vector. .

[0027] Preferably, the acquisition of the fusion feature includes:

[0028] Through two independent linear layers and Will and Projecting onto the same feature dimension yields aligned features. and ;

[0029] Using cross-attention mechanism, features and They act as information sources to recalibrate each other, using global information. To guide local information Generate enhanced local features for:

[0030] ;

[0031] Using local information To enrich overall information Generate enhanced global features for:

[0032] ;

[0033] The original alignment features and The network is spliced ​​together, and the input is a gated network consisting of two fully connected layers. And a gate signal is generated by the Sigmoid function. The gating signal is represented as:

[0034] ;

[0035] Using gating signals The two recalibrated features are dynamically weighted and summed to obtain the final fused features. for:

[0036] ;

[0037] in: This indicates element-wise multiplication.

[0038] Preferably, after acquiring the sMRI data, the process further includes preprocessing the sMRI data, including:

[0039] Remove non-brain tissue from sMRI images to focus on the brain's main structures;

[0040] After removing non-brain tissue, spatial registration is performed to align the individual brain to a standard space, achieving voxel-level spatial coordinate unification.

[0041] Spatial registration is followed by field correction to eliminate image intensity inhomogeneity.

[0042] This invention also provides a brain age prediction system, comprising:

[0043] The imaging module is used to acquire sMRI data of the target subject.

[0044] The feature extraction module is used to extract local detailed features that characterize the boundary texture of the cerebral cortex and the morphological features of brain tissue from sMRI images.

[0045] The sMRI data of nuclear magnetic resonance imaging (MRI) images were sliced ​​into image sequences, and features were extracted from each image sequence under different feature channels. During feature extraction, the dependencies between local regions of different images were obtained based on a windowed multi-head self-attention mechanism, and the importance weights of different feature channels were generated during global spatial feature extraction based on a channel self-attention mechanism. Based on the dependencies and the importance weights of different feature channels, the features extracted from different feature channels were weighted and fused to obtain global detailed features.

[0046] The cross-fusion module is used to perform attention calculation based on the cross-attention mechanism, using global detail features as keys to obtain global attention and weighted fusion to obtain global detail features rich in local features, and using local detail features as queries to perform attention calculation, obtaining local attention and weighted fusion to obtain local detail features enhanced in local features.

[0047] The original global detail features and local detail features are concatenated to obtain concatenated features. The concatenated features are then passed through a fully connected layer and a sigmoid activation function to obtain dynamic gating weights. Based on the dynamic gating weights, the global detail features rich in local features and the local detail features enhanced by local features are dynamically weighted and fused to obtain fused features. The brain age of the target measurer is predicted based on the fused features.

[0048] This invention also provides an electronic device, including a memory and a processor;

[0049] The memory is used to store computer programs;

[0050] When the processor executes the computer program stored in the memory, it implements the steps of the brain age prediction method described above.

[0051] This invention also provides a computer-readable storage medium for storing a computer program, which, when executed by a processor, implements the steps of a brain age prediction method as described above.

[0052] This invention provides a brain age prediction method, system, device, and medium, which have the following advantages compared with the prior art:

[0053] This invention first accurately captures precise local detail features representing the cerebral cortex and brain tissue. Then, while extracting features under different feature channels, it acquires the dependencies between different local regions of the image based on a windowed multi-head self-attention mechanism. It also introduces a channel attention mechanism, learning the importance weights of different feature channels and applying these weights to feature fusion to obtain accurate global detail features. This extraction process simultaneously focuses on what kind and what location of information is more important, accurately capturing long-distance dependencies and subtle structural features. Then, based on a cross-attention mechanism, it guides the extracted global detail features to focus on regions rich in local features, and guides the extracted local detail features to focus on the most relevant local regions, to uncover deeper levels of detail features. It then uses the original global and local detail features to set dynamic gating weights, dynamically adjusting the contribution ratio of local and global contextual detail features in the final decision, and fusing them into the final feature. This achieves a deep and dynamic fusion of two heterogeneous features, thereby gaining a deeper understanding of local and global detail features, ultimately enabling highly accurate brain age prediction. Attached Figure Description

[0054] Figure 1 This is a schematic diagram of the overall process of a brain age prediction method provided in an embodiment of the present invention;

[0055] Figure 2 This is a schematic diagram of a lightweight 3D-CNN module structure for a brain age prediction method provided in an embodiment of the present invention;

[0056] Figure 3 This is a schematic diagram of the structure of an improved Swing Transformer module for a brain age prediction method provided in an embodiment of the present invention;

[0057] Figure 4 This is a schematic diagram of the cross-attention gating fusion module structure of a brain age prediction method provided in an embodiment of the present invention. Detailed Implementation

[0058] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0059] Currently, the main methods used in brain age prediction include 3D-CNN-based methods, Transformer-based methods, pure 3D convolutional neural network (3D-CNN) methods, and pure visual Transformer (VisionTransformer) methods; specifically:

[0060] 1. A method based on 3D-CNN.

[0061] With the development of deep learning technology, predicting brain age directly from sMRI three-dimensional images end-to-end using three-dimensional convolutional neural networks (3D-CNN) has become the most mainstream technical approach. Compared with traditional methods based on handcrafted features such as region volume or cortical thickness, 3D-CNN can automatically learn complex, high-dimensional morphological and textural features related to age in images. Classic 3D-CNN architectures (such as 3D-ResNet, 3D-VGG, etc.) gradually build a hierarchical abstract representation of brain structure by stacking multiple three-dimensional convolutional layers and pooling layers.

[0062] However, such methods suffer from a fundamental technical bottleneck: the inherent local receptive field of convolution operations. Specifically, a single convolutional kernel can only process a local neighborhood of the input data. Even if multiple layers are stacked to expand the receptive field, it is still essentially a layer-by-layer aggregation of local information. The brain is a highly complex, functionally and structurally interconnected network. Its health depends not only on the microstructure of local brain regions but also on the coordination and connections between different brain regions, especially distant ones. Due to the limitations of its design principles, pure 3D-CNN models are severely inadequate in effectively capturing long-range dependencies between different brain regions. For example, it is difficult for them to directly establish structural association features between the prefrontal cortex and the cerebellum. This neglect of global contextual information prevents the model from learning complete brain organization patterns, thus limiting its predictive accuracy.

[0063] 2. Transformer-based methods.

[0064] To overcome the locality limitation of CNNs, some research has begun to explore applying the Transformer model, which has achieved great success in natural language processing, to visual tasks, namely the Vision Transformer (ViT). The core of the Transformer is the self-attention mechanism, which can calculate the relationship between any two elements in the input sequence, thus possessing a powerful ability to model global information. In brain age prediction tasks, 3D brain images are usually segmented into a series of non-overlapping 3D patches, and these patches are serialized and input into the Transformer encoder. Theoretically, this allows the model to pay attention to all regions of the brain at the same time, thereby capturing long-distance dependencies that CNNs ignore.

[0065] However, the pure Transformer approach also brings new problems when applied to medical imaging, especially sMRI. Its preprocessing method of "slicing" images into blocks destroys the fine spatial continuity at the voxel level. For brain age prediction, local texture details such as fine folds in the cerebral cortex and the sharpness of the gray-white matter boundary are extremely important age-related features. The Transformer model is not as direct and effective as CNN in capturing these high-frequency details. In addition, the Transformer model usually requires a larger dataset for pre-training than CNN to converge to ideal performance, which is also a challenge in the medical field where data acquisition is relatively difficult.

[0066] 3. Pure 3D Convolutional Neural Network (3D-CNN) scheme.

[0067] This approach treats preprocessed sMRI three-dimensional brain images as a whole and directly inputs them into an end-to-end deep 3D-CNN model (such as one based on 3D-ResNet or a simplified 3D-CNN architecture). The model performs sliding calculations on the images through a series of three-dimensional convolutional kernels, extracting hierarchical features from low-level texture to high-level morphology layer by layer, and finally obtains the predicted brain age value through regression of fully connected layers.

[0068] The main drawback of this scheme is its lack of ability to represent global context information.

[0069] (1) This defect stems from the inherent mechanism of convolution operation: the computation range of each convolution kernel is strictly limited to the size of its receptive field, which is essentially a local operation; even if the theoretical receptive field is expanded by increasing the number of network layers, this expansion is indirect and inefficient, and it is impossible to establish a direct and explicit long-distance dependency model between different brain regions.

[0070] (2) Brain aging is not an isolated event, but the result of systemic changes in the entire brain network; for example, hippocampal atrophy may be related to the functional degeneration of the distant entorhinal cortex. Pure 3D-CNN schemes are difficult to capture such cross-brain region structural association information. The features they learn reflect more the morphological changes of local areas, while ignoring the systemic organizational pattern of the brain as a whole; this one-sidedness of information directly leads to the bottleneck of its predictive performance.

[0071] 4. Pure Vision Transformer Solution.

[0072] This approach draws inspiration from natural language processing. First, the 3D brain image is spatially segmented into a series of fixed-size, non-overlapping 3D image patches. These patches are then linearly flattened and positionally encoded to form a sequence that is input into the Transformer encoder. The encoder's self-attention mechanism calculates the association weights between each image patch in the sequence and all other image patches, thereby constructing a global feature representation that is ultimately used for brain age prediction.

[0073] The main drawback of this scheme is its insufficient ability to perceive fine local spatial information.

[0074] (1) This defect stems from its “segmentation first, processing later” working paradigm: the continuous sMRI image is forcibly segmented into discrete image blocks, which itself destroys the spatial continuity at the boundary of the image block; although the self-attention mechanism can establish the relationship between blocks, it lacks a strong inductive bias like convolution operation for the voxel-level fine structure inside the block.

[0075] (2) In sMRI images, high-frequency details such as the complex morphology of cortical sulci, the clarity of the gray-white matter interface, and the texture of small lesions are key biomarkers that are highly correlated with age. The pure Transformer approach, due to its insensitivity to local details, may not be able to make full use of these details, resulting in a lack of precision in feature extraction and thus affecting the final prediction accuracy.

[0076] As can be seen from the above analysis, 3D-CNN is good at extracting local details and spatial textures, but its "field of view" is limited and it lacks a global perspective. Transformer is good at establishing global long-distance dependencies, but its processing method is coarse and it is easy to lose fine local information.

[0077] To address the aforementioned shortcomings of existing technical solutions—namely, the inability of pure 3D-CNN models to effectively capture global long-range dependencies, and the tendency of pure Transformer models to lose detailed local spatial information, resulting in incomplete representation of brain imaging information and thus limiting the accuracy of brain age prediction; for example... Figure 1 As shown, this invention proposes an end-to-end brain age prediction scheme, mainly comprising four core stages: 1. Data preprocessing: The original sMRI brain images undergo a series of standardization processes to eliminate inter-individual differences, obtaining image data with consistent size and spatial orientation; 2. Dual-branch feature extraction: The preprocessed images are simultaneously input into two parallel feature extraction branches: one is a lightweight 3D-CNN module for capturing fine local spatial textures; the other is an improved Swin Transformer module for establishing global long-range dependencies; 3. Feature fusion: The deep features output from the two branches are fed into the core cross-attention gating fusion module of this invention for intelligent information interaction and dynamic fusion; 4. Brain age regression: The fused final features are augmented with residuals and then input into a multilayer perceptron (MLP) regression head to output the final predicted brain age value. Specifically:

[0078] Step 1: Brain image preprocessing.

[0079] To ensure the standardization and consistency of model input, all raw sMRI images undergo standardized preprocessing. This preprocessing workflow is consistent with that of large public datasets such as UK Biobank, and mainly includes: skull stripping to remove non-brain tissue, affine transformation-based spatial registration to MNI152 standard space, and bias field correction. After preprocessing, all brain images are unified into standard three-dimensional data with the same XYZ coordinate dimensions, which are then used as input for the subsequent model.

[0080] Step 2: Parallel dual-branch feature extraction.

[0081] 2.1 Lightweight 3D-CNN module (local feature branch).

[0082] Figure 2This paper showcases the architecture of a lightweight 3D-CNN module designed for efficient extraction of local spatial details from sMRI images, such as high-frequency features like cortical texture, gray-white matter boundary morphology, and local tissue density. The module is a feedforward network without residual connections, consisting of six cascaded basic units. Each basic unit contains a 3x3x3 convolutional layer and a 2x2x2 max-pooling layer. This streamlined design ensures effective extraction of local features while reducing the number of parameters to half that of a standard 3D-CNN network with the same number of layers and channels, significantly lowering computational overhead. The network terminates with a Global Average Pooling (GAP) layer, aggregating the final 3D feature map into a one-dimensional deep convolutional feature vector. Mathematical model: For the first The 3D convolution operation of the layer outputs a feature map. any voxel on The value can be represented as:

[0083] .

[0084] in: It is the input feature map; and These are the kernel weights and biases of this layer. It is an activation function.

[0085] 2.2 Improved Swing Transformer module (global feature branch).

[0086] Figure 3 This paper demonstrates the architecture of the improved Swin Transformer module used in this invention. This module is designed to capture long-range dependencies between different brain regions, i.e., global contextual information. To accommodate the input of the 2D Transformer, the 3D brain image is first segmented along the Z-axis (sagittal or coronal plane is also acceptable) into a sequence of N 2D image slices. This module is based on the Swin Transformer architecture, and its core is the parallel use of Window-based Multi-head Self-Attention (W-MSA) and Channel-based Multi-head Self-Attention (CBA). The key improvement of this invention lies in the innovative introduction of a channel attention mechanism on top of the original self-attention computation. Specifically, in each Transformer block, in addition to calculating attention in the spatial dimension, a channel attention module is connected in parallel. This module learns the importance weights of different feature channels and applies them to the feature map. This allows the model to not only focus on "where" information is more important (spatial attention) but also on "what" information is more important (channel attention). For an input sequence... ( For sequence length, (where the feature dimension is 1), standard self-attention computation can be expressed as:

[0087] .

[0088] in: These are the query, key, and value matrices obtained through linear transformation. The Swin Transformer restricts this to non-overlapping local windows (W-MSA) and achieves cross-window information interaction through window shifting (SW-MSA). The improvement of this invention is to introduce an additional channel attention weight. The module performs element-wise multiplication with the output features to enhance feature representation; the final output of this module is a one-dimensional deep long-range feature vector. .

[0089] Step 3: Cross-attention gating fusion

[0090] Figure 4 This invention showcases the core innovation of its design: a cross-attention gating fusion module; this module aims to achieve... and Deep and dynamic fusion of two heterogeneous features, rather than simple concatenation or addition; allowing local detail features and global context features to guide and enhance each other, and adaptively learning their contribution weights in the final prediction; fusion process and mathematical model:

[0091] 1. Feature alignment: First, through two independent linear layers ( and )Will and Projecting onto the same feature dimension yields aligned features. and .

[0092] 2. Cross-attention relabeling: Utilizing the cross-attention mechanism, two features act as "information sources" to relabel each other; using global information... To guide local information Generate enhanced local features for:

[0093] .

[0094] Using local information To enrich overall information Generate enhanced global features for:

[0095] .

[0096] 3. Gating signal generation: This involves generating the original alignment features. and Concatenate the layers, then input a gated network consisting of two fully connected layers. And a gate signal is generated by the Sigmoid function. Its value ranges between (0, 1), and its gate signal is represented as:

[0097] .

[0098] 4. Dynamic weighted fusion: utilizing gating signals The two recalibrated features are dynamically weighted and summed to obtain the final fused features. for:

[0099] .

[0100] in: This indicates element-wise multiplication; this gating mechanism allows the model to adaptively determine, based on the input data, whether to focus more on local details enhanced by global information or on global context enriched by local information in each feature dimension.

[0101] Step 4: Brain age regression.

[0102] To prevent information loss during deep fusion, the fusion features will be... Compared with the original deep features and The features are then aggregated again using residual connections. Finally, the enhanced final feature vector is input into a regression head composed of a multilayer perceptron (MLP) to output the final predicted brain age.

[0103] This invention aims to overcome the limitations of a single architecture: through a parallel dual-branch structure, it utilizes a lightweight 3D-CNN and an improved Swing Transformer respectively to specifically and efficiently extract the local spatial texture and global long-range dependencies of the brain; it aims to achieve deep feature synergy: a novel cross-attention gating fusion module is designed to replace simple feature splicing or addition. This module can guide two heterogeneous features to perform intelligent information interaction and dynamic weighting, achieving complementary advantages and redundancy suppression; it aims to improve prediction accuracy: through the above design, features with richer information and stronger representational capabilities are ultimately generated, thereby significantly improving the accuracy of brain age prediction and making it more reliable than existing single-technology solutions.

[0104] Experimental validation of this invention on large public datasets shows that, compared to the mean absolute error (MAE) of approximately 3.5 years for existing mainstream solutions, the method of this invention can successfully reduce the MAE to 2.9 years. This improvement in accuracy stems from the invention's ability to generate more comprehensive and robust brain feature representations. By fusing local details with global context, the model can capture complex age-related structural patterns that are ignored by single-architecture models, thereby making judgments closer to the physiological true age. This provides more reliable biomarkers for clinical diagnosis and health assessment, significantly improving prediction accuracy and making the results more reliable.

[0105] The dual-branch parallel structure and core cross-attention gating fusion module designed in this invention overcome the problem of "one-sidedness" in existing technologies. It does not simply splice two features together, but achieves intelligent information interaction through a cross-attention mechanism—allowing global information to guide the model to focus on key local details, while allowing local details to provide concrete evidence for abstract global relationships. The gating mechanism further realizes dynamic and adaptive fusion. This synergistic effect of 1+1>2 makes the final fused features more representative and discriminative than features from any single source, resulting in stronger feature representation capabilities and more comprehensive information utilization.

[0106] While pursuing high accuracy, this invention fully considers the practicality of the model. Its 3D-CNN branch adopts a lightweight design, with only half the number of parameters as a standard network, effectively reducing the overall computational burden and storage requirements of the model. This makes the model of this invention have a better performance-to-power ratio compared to some heavy networks that also pursue high performance but have bloated structures (such as deep 3D-ResNet). It is also easier to train and deploy on conventional computing workstations, enhancing the clinical translation potential of the technical solution and achieving a better balance between computational efficiency and model performance.

[0107] This invention employs the guiding role of cross-attention, a mechanism that uses global contextual features as a "compass" to guide 3D-CNN features to focus on local regions most relevant to the global aging pattern, avoiding local branches getting caught up in irrelevant textures. Conversely, local detail features also provide concrete anatomical basis for abstract global dependencies, making the Transformer's learning process more interpretable.

[0108] The introduction of gating signals in this invention enables the model to dynamically adjust the contribution ratio of local details and global context in the final decision based on the unique characteristics of each input sample. This data-driven adaptive capability is not available in simple fusion methods (such as concatenation and addition), which greatly improves the discriminative robustness of features.

[0109] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A brain age prediction method, characterized in that, Includes the following steps: Acquire sMRI data of the target subject; Extract local detailed features from sMRI data that characterize the boundary texture of the cerebral cortex and the morphological features of brain tissue; The sMRI data of nuclear magnetic resonance imaging (MRI) images were sliced ​​into image sequences, and features were extracted from each image sequence under different feature channels. During feature extraction, the dependencies between local regions of different images were obtained based on a windowed multi-head self-attention mechanism, and the importance weights of different feature channels were generated during global spatial feature extraction based on a channel self-attention mechanism. Based on the dependencies and the importance weights of different feature channels, the features extracted from different feature channels were weighted and fused to obtain global detailed features. Based on the cross-attention mechanism, global detail features are used as keys for attention calculation, global attention is obtained and weighted fusion is performed to obtain global detail features rich in local features, and local detail features are used as queries for attention calculation, local attention is obtained and weighted fusion is performed to obtain local detail features enhanced in local features. The original global detail features and local detail features are concatenated to obtain concatenated features. The concatenated features are then passed through a fully connected layer and a Sigmoid activation function to obtain dynamic gating weights. Based on the dynamic gating weights, the global detail features with rich local features and the local detail features with enhanced local features are dynamically weighted and fused to obtain fused features. The brain age of the target measurer is predicted based on the fused features. The acquisition of the local detail features includes: Feature extraction is performed using a 3x3x3 convolutional layer to obtain a 3D feature map. This 3D feature map is then passed through a 2x2x2 max pooling layer to aggregate it into a 1D deep convolutional feature vector. ; Among them, for the third-dimensional convolutional layer The 3D convolution operation of the layer outputs a feature map. any voxel on The value is represented as: ； in: Indicates the input feature map; and This indicates the kernel weights and biases of this layer; Indicates the activation function; The acquisition of the global detailed features includes: For an input image sequence ,in For sequence length, As the feature dimension, self-attention computation is represented as: ； in: These represent the query, key, and value matrices obtained through linear transformation, respectively. Windowed multi-head self-attention restricts it to non-overlapping local windows and enables cross-window information interaction through window shifting SW-MSA; Based on information interaction, a channel attention weight is introduced. Element-wise multiplication with the output features is performed to enhance the feature representation, resulting in a one-dimensional deep long-range feature vector. .

2. The brain age prediction method according to claim 1, characterized in that, The fusion features are obtained through a brain age prediction model, which includes a cascaded lightweight 3D-CNN module, a Swing Transformer module, and a cross-attention gating fusion module. The lightweight 3D-CNN module consists of 6 basic units connected in series. Each basic unit contains a 3D convolutional layer with a kernel size of 3x3x3 and a max pooling layer with a kernel size of 2x2x2, which is used to extract local detail features. The Swin Transformer module is based on the Swin Transformer architecture and introduces a channel self-attention mechanism in parallel in each windowed multi-head self-attention block to extract global detail features. The cross-attention gating fusion module includes cascaded linear layers, a cross-attention mechanism, and a fully connected layer. The activation function is the Sigmoid function, which is used to extract global detail features rich in local features and local detail features enhanced by local features, and to dynamically weight and fuse the global detail features rich in local features and the local detail features enhanced by local features.

3. The brain age prediction method according to claim 2, characterized in that, The acquisition of the fusion features includes: Through two independent linear layers and Will and Projecting onto the same feature dimension yields aligned features. and ; Using cross-attention mechanism, features and They act as information sources to recalibrate each other, using global information. To guide local information Generate enhanced local features for: ； Using local information To enrich overall information Generate enhanced global features for: ； The original alignment features and The network is spliced ​​together, and the input is a gated network consisting of two fully connected layers. And a gate signal is generated by the Sigmoid function. The gating signal is represented as: ； Using gating signals The two recalibrated features are dynamically weighted and summed to obtain the final fused features. for: ； in: This indicates element-wise multiplication.

4. The brain age prediction method according to claim 1, characterized in that, After acquisition, the sMRI data of the nuclear magnetic resonance imaging (sMRI) images also includes preprocessing, including: Remove non-brain tissue from sMRI images to focus on the brain's main structures; After removing non-brain tissue, spatial registration is performed to align the individual brain to a standard space, achieving voxel-level spatial coordinate unification. Spatial registration is followed by field correction to eliminate image intensity inhomogeneity.

5. A brain age prediction system, characterized in that, include: The imaging module is used to acquire sMRI data of the target subject. The feature extraction module is used to extract local detailed features that characterize the boundary texture of the cerebral cortex and the morphological features of brain tissue from sMRI images. The sMRI data of nuclear magnetic resonance imaging (MRI) images were sliced ​​into image sequences, and features were extracted from each image sequence under different feature channels. During feature extraction, the dependencies between local regions of different images were obtained based on a windowed multi-head self-attention mechanism, and the importance weights of different feature channels were generated during global spatial feature extraction based on a channel self-attention mechanism. Based on the dependencies and the importance weights of different feature channels, the features extracted from different feature channels were weighted and fused to obtain global detailed features. The cross-fusion module is used to perform attention calculation based on the cross-attention mechanism, using global detail features as keys to obtain global attention and weighted fusion to obtain global detail features rich in local features, and using local detail features as queries to perform attention calculation, obtaining local attention and weighted fusion to obtain local detail features enhanced in local features. The original global detail features and local detail features are concatenated to obtain concatenated features. The concatenated features are then passed through a fully connected layer and a Sigmoid activation function to obtain dynamic gating weights. Based on the dynamic gating weights, the global detail features with rich local features and the local detail features with enhanced local features are dynamically weighted and fused to obtain fused features. The brain age of the target measurer is predicted based on the fused features. The acquisition of the local detail features includes: Feature extraction is performed using a 3x3x3 convolutional layer to obtain a 3D feature map. This 3D feature map is then passed through a 2x2x2 max pooling layer to aggregate it into a 1D deep convolutional feature vector. ; Among them, for the third-dimensional convolutional layer The 3D convolution operation of the layer outputs a feature map. any voxel on The value is represented as: ； in: Indicates the input feature map; and This indicates the kernel weights and biases of this layer; Indicates the activation function; The acquisition of the global detailed features includes: For an input image sequence ,in For sequence length, As the feature dimension, self-attention computation is represented as: ； in: These represent the query, key, and value matrices obtained through linear transformation, respectively. Windowed multi-head self-attention restricts it to non-overlapping local windows and enables cross-window information interaction through window shifting SW-MSA; Based on information interaction, a channel attention weight is introduced. Element-wise multiplication with the output features is performed to enhance the feature representation, resulting in a one-dimensional deep long-range feature vector. .

6. An electronic device, characterized in that, include: Memory and processor; The memory is used to store computer programs; When the processor executes the computer program stored in the memory, it implements the steps of the brain age prediction method as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, Used to store a computer program, which, when executed by a processor, implements the steps of a brain age prediction method as described in any one of claims 1 to 4.

Images