Brain age prediction method, system, equipment and medium

Abstract

The invention discloses a brain age prediction method, system and device and a medium, and relates to the technical field of biomedical image analys.The method comprises the steps that firstly, precise local detail features representing the cerebral cortex and brain tissue are precisely captured, then the dependency relationship between local areas of different images is obtained based on a windowed multi-head self-attention mechanism, and the brain age prediction result is obtained; a channel attention mechanism is introduced, importance weights of different feature channels are learned, and the importance weights are applied to feature fusion, so that long-distance dependency relationships and fine structure features are accurately captured; and then based on a cross attention mechanism, guiding global detail features to focus on a region with rich local features and guiding the local detail features to focus on a most relevant local region so as to mine deeper detail features, and finally dynamically adjusting the contribution proportion of the local detail features and global context detail features in a final decision. And the features are fused into final features so as to perceive local detail features and global detail features in a deeper level.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of biomedical image analysis, in particular to a brain age prediction method, system, device and medium. BACKGROUND

[0002] "Brain Age" is a biomarker in the field of neuroscience and clinical medicine that has attracted much attention in recent years. It uses a machine learning model established based on 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 actual age of the individual is called "Brain Age Gap" (BAG). Numerous studies have shown that a positive brain age gap (i.e., a predicted brain age 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 disorders, such as Alzheimer's disease, Parkinson's disease, schizophrenia, depression, etc. Therefore, developing a high-precision brain age prediction technology not only has great scientific value in understanding the health trajectory of an individual's brain and achieving 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] At present, the brain age prediction methods based on deep learning mainly include methods based on three-dimensional convolutional neural networks (3D-CNN) and methods based on visual Transformers (ViT). The method based on three-dimensional convolutional neural networks (3D-CNN) can effectively extract fine structural features (such as gray-white matter boundaries, cortical folds, etc.) in sMRI through local sliding of convolution kernels. However, due to the limitations of the local receptive field of convolution operations, it is difficult to effectively capture the long-range dependencies between different regions of the brain, and the ability to perceive global details is weak when perceiving local details. The method based on visual Transformers (ViT) can capture global dependencies through self-attention mechanisms, but the way of cutting and processing 3D images destroys the spatial continuity at the voxel level, and the ability to perceive local details is weak when modeling global information.

[0004] Therefore, the current brain age prediction methods based on deep learning are difficult to simultaneously capture precise long-range dependencies and fine structural features between different regions of the brain, resulting in difficulty in simultaneously perceiving local and global detail features at a deeper level. SUMMARY

[0005] The present application provides a brain age prediction method, system, device and medium, which can solve the problems in the prior art.

[0006] The present application provides a brain age prediction method, comprising the following steps: obtaining nuclear magnetic resonance image sMRI data of a target measurer; extracting local detail features representing texture boundary of cerebral cortex edge and representing morphological features of brain tissue in the nuclear magnetic resonance image sMRI data; slicing the nuclear magnetic resonance image sMRI data into image sequences, performing feature extraction on each image sequence under different feature channels, obtaining dependency relationships between different local regions of the image based on windowed multi-head self-attention mechanism during feature extraction, and generating importance weights of different feature channels during global spatial feature extraction based on channel self-attention mechanism; weighting and fusing the features extracted by different feature channels according to the dependency relationships and the importance weights of different feature channels to obtain global detail features; based on the cross-attention mechanism, performing attention calculation on the global detail features as keys to obtain global attention and weighting fusion to obtain local feature-rich global detail features, and performing attention calculation on the local detail features as queries to obtain local attention and weighting fusion to obtain local feature-enhanced local detail features; concatenating the original global detail features and the local detail features to obtain concatenated features, and obtaining dynamic gating weights by passing the concatenated features through a fully connected layer and a Sigmoid activation function; based on the dynamic gating weights, dynamically weighting and fusing the local feature-rich global detail features and the local feature-enhanced local detail features to obtain fusion features, and predicting the brain age of the target measurer based on the fusion features.

[0007] Preferably, the fusion features are obtained by a brain age prediction model, and the brain age prediction model comprises a cascaded lightweight 3D-CNN module, a Swin Transformer module and a cross-attention gating fusion module. The lightweight 3D-CNN module is composed of 6 basic units connected in series, each basic unit containing a three-dimensional convolution layer with a convolution kernel size of 3x3x3 and a maximum pooling layer with a pooling kernel size of 2x2x2, for extracting local detail features. The Swin Transformer module is based on the Swin Transformer architecture, and the channel self-attention mechanism is introduced in parallel in each windowed multi-head self-attention block, for extracting global detail features. The cross-attention gating fusion module comprises a cascaded linear layer, a cross-attention mechanism and a fully connected layer, wherein the activation function adopts a Sigmoid function, for extracting local feature-rich global detail features and local feature-enhanced local detail features, and dynamically weighting and fusing the local feature-rich global detail features and the local feature-enhanced local detail features.

[0008] Preferably, 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; This represents the activation function.

[0009] Preferably, 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. .

[0010] Preferably, the acquisition of the fusion feature 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 To: ; Use local information to enrich global information , generate enhanced global features To: ; Splice the original alignment features And Input a gating network composed of two fully connected layers , and generate a gating signal through the Sigmoid function , the gating signal is represented as: ; Use the gating signal To dynamically weight and sum the two features after re-calibration, obtain the final fusion features To: ; Where: Indicates element-wise multiplication.

[0011] Preferably, the nuclear magnetic resonance image sMRI data after acquisition, further comprising preprocessing the nuclear magnetic resonance image sMRI data, comprising: Remove non-brain tissue in the nuclear magnetic resonance image sMRI to focus on the brain structure; After removing the non-brain tissue, spatial registration is performed to align the individual brain to the standard space, and the voxel level spatial coordinates are unified; After spatial registration, perform field correction to eliminate image intensity non-uniformity.

[0012] The embodiment of the application also provides a brain age prediction system, comprising: An image module for acquiring nuclear magnetic resonance image sMRI data of a target measurer; A feature extraction module for extracting local detail features representing the texture boundary of the cerebral cortex edge and representing the morphological features of the brain tissue in the nuclear magnetic resonance image sMRI data; The nuclear magnetic resonance image sMRI data is sliced into image sequences, and feature extraction is performed on each image sequence under different feature channels. At the same time of feature extraction, the dependency relationship between different local regions of the image is obtained based on the windowed multi-head self-attention mechanism, and the importance weight of different feature channels is generated based on the channel self-attention mechanism during global spatial feature extraction. According to the dependency relationship and the importance weight of different feature channels, the features extracted by different feature channels are weighted and fused to obtain global detail features; The cross fusion module is used for attention calculation of the global detail feature as a key value based on a cross attention mechanism, obtaining global attention and weighted fusion to obtain a global detail feature rich in local features, and attention calculation of the local detail feature as a query, obtaining local attention and weighted fusion to obtain a local detail feature enhanced in local features; The original global detail feature and the local detail feature are spliced to obtain spliced features, the spliced features pass through a full connection layer and a Sigmoid activation function to obtain a dynamic gating weight, and the global detail feature rich in local features and the local detail feature enhanced in local features are dynamically weighted and fused based on the dynamic gating weight to obtain a fusion feature, and the brain age of the target measurer is predicted based on the fusion feature.

[0013] The embodiment of the present application also provides an electronic device comprising a memory and a processor. The memory is used for storing a computer program. The processor is used for executing the computer program stored in the memory to realize the steps of the brain age prediction method.

[0014] The embodiment of the present application also provides a computer readable storage medium for storing a computer program, and the computer program is executed by the processor to realize the steps of the brain age prediction method.

[0015] The embodiment of the present application provides a brain age prediction method, system, device and medium, and has the following beneficial effects compared with the prior art. The embodiment of the present application firstly accurately captures the local detail features representing the brain cortex and brain tissue, then simultaneously extracts features under different feature channels, obtains the dependency relationship between different local regions of the image based on the windowed multi-head self-attention mechanism, introduces the channel attention mechanism, learns the importance weight of different feature channels, and applies the importance weight to feature fusion to obtain accurate global detail features, and the extraction process can simultaneously pay attention to what information and what position is more important to accurately capture long-distance dependency relationship and subtle structure features; then based on the cross attention mechanism, the extracted global detail features are guided to focus on the local feature rich region, the extracted local detail features are guided to focus on the most relevant local region to mine deeper detail features, and the dynamic gating weight is set based on the original global detail feature and the local detail feature, the contribution proportion of the local detail feature and the global context detail feature in the final decision is dynamically adjusted, and the final feature is fused to realize the deep and dynamic fusion of the two heterogeneous features, so that the local detail feature and the global detail feature are perceived in a deeper level, and finally the brain age is predicted with high accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1A schematic diagram of the overall process of a brain age prediction method provided for an embodiment of the present application is shown in FIG. 1. Figure 2 A schematic diagram of a lightweight 3D-CNN module structure of a brain age prediction method provided for an embodiment of the present application is shown in FIG. 2. Figure 3 A schematic diagram of an improved Swin Transformer module structure of a brain age prediction method provided for an embodiment of the present application is shown in FIG. 3. Figure 4 A schematic diagram of a cross-attention gate fusion module structure of a brain age prediction method provided for an embodiment of the present application is shown in FIG. 4. DETAILED DESCRIPTION

[0017] In order to make the above-mentioned objects, features and advantages of the present application more apparent and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the accompanying drawings. In the following description, a large number of specific details are set forth in order to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the scope of the present application, so the present application is not limited to the specific embodiments disclosed below.

[0018] Currently, the means used in brain age prediction mainly include methods based on 3D-CNN, methods based on Transformer, pure three-dimensional convolutional neural network (3D-CNN) means, and pure visual Transformer (Vision Transformer) means; specifically: 1. Method based on 3D-CNN.

[0019] With the development of deep learning technology, using three-dimensional convolutional neural network (3D-CNN) to directly predict brain age from sMRI three-dimensional images end-to-end has become the most mainstream technical route at present; compared with traditional methods based on regional volume or cortical thickness and other manual features, 3D-CNN can automatically learn complex, high-dimensional morphological and textural features related to age in images; classic 3D-CNN architecture (such as 3D-ResNet, 3D-VGG, etc.) gradually builds a hierarchical abstract representation of brain structure by stacking multiple three-dimensional convolutional layers and pooling layers.

[0020] However, such methods have a fundamental technical bottleneck: the local receptive field inherent in convolution operations; specifically, a single convolution kernel can only process a local neighborhood in the input data, and even if the receptive field is expanded by stacking multiple layers, it is essentially still a layer-by-layer aggregation of local information; the brain is a highly complex, functionally and structurally integrated network, and its health status depends not only on the microstructure of local brain regions, but also on the coordination and association between different brain regions, especially long-distance brain regions; due to the limitations of its design principles, pure 3D-CNN models are severely inadequate in effectively capturing long-range dependencies between different regions of the brain; for example, it is difficult for them to directly establish structural association features between the prefrontal cortex and the cerebellum; this neglect of global context information prevents the model from learning the complete brain organization pattern, thereby limiting its prediction accuracy ceiling.

[0021] 2. Transformer-based methods.

[0022] To overcome the local limitations of CNNs, some studies have begun to explore the application of the Transformer model, which has achieved great success in natural language processing, to visual tasks, i.e., Vision Transformer (ViT); the core of Transformer is the self-attention mechanism, which can calculate the mutual relationship between any two elements in the input sequence, thus inherently possessing strong global information modeling capabilities; in the brain age prediction task, 3D brain images are typically divided into a series of non-overlapping 3D patches, and these patch sequences are input into the Transformer encoder; in theory, this allows the model to simultaneously focus on all regions of the brain, thereby capturing long-range dependencies that CNNs miss.

[0023] However, the pure Transformer approach also presents new problems when applied to medical images, particularly sMRI; the preprocessing method of "patching" the image disrupts the fine spatial continuity at the voxel level; for brain age prediction, the fine texture details of the cerebral cortex, such as the sharpness of the gray-white matter junction, are extremely important age-related features; Transformer models are not as direct and effective as CNNs in capturing these high-frequency detail information; in addition, Transformer models typically require larger-scale datasets for pre-training to converge to ideal performance, which is also a challenge in the medical field where data acquisition is relatively difficult.

[0024] 3. Pure three-dimensional convolutional neural network (3D-CNN) scheme.

[0025] This scheme takes the pre-processed sMRI three-dimensional brain image as a whole and directly inputs it into an end-to-end deep 3D-CNN model (such as based on 3D-ResNet or simplified 3D-CNN architecture); the model performs sliding calculation on the image through a series of three-dimensional convolution kernels, extracts hierarchical features from low-level texture to high-level morphology layer by layer, and finally regresses the predicted brain age value through the fully connected layer.

[0026] The main defect of this scheme is the lack of representation ability of global context information: (1) This defect is due to the inherent mechanism of convolution operation: the calculation range of each convolution kernel is strictly limited within its receptive field size, which is essentially a local operation; even if the network depth is deepened to expand the theoretical receptive field, this expansion is indirect and inefficient, and cannot establish a direct and explicit long-distance dependency model between different brain regions.

[0027] (2) Brain aging is not an isolated event, but the result of systematic changes in the entire brain network; for example, the atrophy of the hippocampus may be related to the functional degradation of the distant entorhinal cortex. The pure 3D-CNN scheme is difficult to capture such cross-brain region structural correlation information, and the features learned by it more reflect the morphological changes of local regions, ignoring the systematic organizational patterns of the brain as a whole; this one-sidedness of information directly leads to the bottleneck of its prediction performance.

[0028] 4. Pure vision Transformer (Vision Transformer) scheme.

[0029] This scheme draws on the ideas of natural language processing, first dividing the 3D brain image into a series of fixed-size, non-overlapping 3D image blocks (Patches) in space; these image blocks are linearly flattened and added with position encoding to form a sequence input into the Transformer encoder; the self-attention mechanism inside the encoder will calculate the correlation weights between each image block in the sequence and all other image blocks, thereby constructing a global feature representation and ultimately used for brain age prediction.

[0030] The main defect of this scheme is the lack of ability to perceive fine local spatial information: (1) The defect is derived from its "segmentation first, then processing" working paradigm: the continuous sMRI image is forced to be segmented into discrete image blocks, which itself destroys the spatial continuity at the block boundary; although the self-attention mechanism can establish the relationship between blocks, it lacks the strong inductive bias of convolution operation for the fine structure of the voxel level within the block.

[0031] (2) In sMRI images, high-frequency detailed information such as complex morphology of cortical sulci, clarity of gray-white matter interface, and texture of small lesions are key biomarkers highly related to age; the pure Transformer scheme may not fully utilize these information due to its insensitivity to local details, resulting in a lack of precision in feature extraction, which affects the final prediction accuracy From the above analysis, 3D-CNN is good at extracting local detail spatial texture, but its "field of view" is limited and lacks global perspective, while Transformer is good at establishing global long-distance dependence, but its processing method is rough and easy to lose fine local information.

[0032] In view of the above defects existing in the prior art: that is, the pure 3D-CNN model cannot effectively capture the global long-distance dependence relationship, and the pure Transformer model is easy to lose the local fine spatial information, resulting in that the brain image information representation is not comprehensive, and then the brain age prediction accuracy is limited; as shown in Figure 1 As shown in the formula, the present application proposes an end-to-end brain age prediction scheme, mainly including four core stages: 1. Data preprocessing: a series of standardization processing is performed on the original sMRI brain image to eliminate individual differences, and image data with uniform size and spatial orientation is obtained; 2. Dual-branch feature extraction: the preprocessed image is input into two parallel feature extraction branches: one is a lightweight 3D-CNN module for capturing fine local spatial texture; the other is an improved Swin Transformer module for establishing global long-distance dependence relationship; 3. Feature fusion: the deep features output by the two branches are input into the core cross-attention gate fusion module of the present application for intelligent information interaction and dynamic fusion; 4. Brain age regression: after residual enhancement, the final feature after fusion is input into a multi-layer perceptron (MLP) regression head, and the final predicted brain age value is output. Specifically:

[0033] Step one: brain image preprocessing.

[0034] To ensure the standardization and consistency of model input, all original sMRI images need to be standardized and pre-processed; the pre-processing procedure is consistent with the processing method of large public data sets such as UK Biobank, mainly including: removing non-brain tissue (Skull Stripping), registration to MNI152 space based on affine transformation, and bias field correction. After pre-processing, all brain images are unified into standard three-dimensional data with the same XYZ coordinate size as the input of the subsequent model.

[0035] Step two: parallel double-branch feature extraction.

[0036] 2.1 Lightweight 3D-CNN module (local feature branch).

[0037] Figure 2 The architecture of the lightweight 3D-CNN module designed by the application is shown, which is dedicated to efficiently extracting local spatial detail information in sMRI images, such as cortical texture, gray-white matter boundary morphology, local tissue density, and other high-frequency features; this module is a feedforward network without residual connection, composed of 6 basic units in series; each basic unit contains a three-dimensional convolution layer with a convolution kernel size of 3x3x3, and a maximum pooling layer with a pooling kernel size of 2x2x2. This simplified design not only ensures the effective extraction of local features, but also reduces the parameter quantity to half of that of a standard 3D-CNN network with the same number of layers and channels, significantly reducing the computational overhead; the end of the network is aggregated into a one-dimensional deep convolution feature vector by a global average pooling (GAP) layer ; mathematical model: for the third layer of three-dimensional convolution operation, the value of any voxel on the output feature map can be represented as:

[0038] .

[0039] Where: is the input feature map; and are the convolution kernel weights and bias of the layer, is the activation function.

[0040] 2.2 Improved Swin Transformer module (global feature branch).

[0041] Figure 3The architecture of the improved Swin Transformer module used in the application is shown, which is designed to capture long-distance dependency between different regions of the brain, i.e. global context information; in order to adapt to the input of the 2D Transformer, the 3D brain image is first divided along the Z axis (sagittal or coronal) into a sequence composed of N 2D image slices ; the module is based on the Swin Transformer architecture, the core of which 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 the application is the innovative introduction of channel attention mechanism on the basis of the original self-attention calculation; specifically, in each Transformer block, in addition to calculating the attention in the spatial dimension, a channel attention module is also connected in parallel, which learns the importance weight of different feature channels and applies it to the feature map; this makes the model not only pay attention to where the information is more important (spatial attention), but also pay attention to what kind of information is more important (channel attention); for an input sequence ( is the sequence length, is the feature dimension), the standard self-attention calculation can be represented as: .

[0042] where: are the query, key and value matrices obtained by linear transformation. Swin Transformer limits it within a non-overlapping local window (W-MSA) and realizes information interaction across windows through window shifting (SW-MSA). The improvement of the application is to additionally introduce a channel attention weight and element-wise multiply with the output feature to enhance the feature representation; the module finally outputs a one-dimensional deep long-distance feature vector .

[0043] Step three: cross-attention gating fusion Figure 4 The core innovation designed by the application is shown: the cross-attention gating fusion module; the module is designed to realize and two kinds of heterogeneous features are not simply spliced or added; let the local detail features and global context features guide each other, enhance each other, and adaptively learn the contribution weight of both in the final prediction; the fusion process and mathematical model are: 1. Feature alignment: First, project the two independent linear layers (L1, L2) to the same feature dimension, get the aligned features and . and . and .

[0044] 2. Cross-attention re-labeling: Use cross-attention mechanism to let the two features re-label each other; use global information to guide local information , generate enhanced local features : .

[0045] Use local information to enrich global information , generate enhanced global features : .

[0046] 3. Gating signal generation: Concatenate the original aligned features and , then input a gating network composed of two fully connected layers , and generate a gating signal whose value range is between (0, 1), the gating signal is expressed as: .

[0047] 4. Dynamic weighted fusion: Use the gating signal to dynamically weight the sum of the two re-labeled features, get the final fusion feature : .

[0048] Where: represents element-wise multiplication; this gating mechanism allows the model to adaptively decide, according to the input data, whether to focus more on the local details enhanced by global information or on the global context enriched by local information in each feature dimension.

[0049] Step four: Brain age regression.

[0050] To prevent information loss during deep fusion, fuse the fusion feature with the original deep features and The residual connection is used to aggregate again; finally, the enhanced final feature vector is input into a regression head composed of a multi-layer perceptron (MLP) to output the final predicted brain age.

[0051] The present application aims to overcome the limitations of a single architecture: through a parallel double-branch structure, respectively using a lightweight 3D-CNN and an improved Swin Transformer, to specifically and efficiently extract the local spatial texture and global long-distance dependency of the brain; aims to realize the depth coordination of features: design a novel cross-attention gated fusion module to replace simple feature splicing or addition, which can guide the intelligent information interaction and dynamic weighting of two heterogeneous features, realize complementary advantages and redundancy suppression; aims to improve the accuracy of prediction: through the above design, the final generated feature is more informative and has stronger representation ability, thereby significantly improving the accuracy of brain age prediction, making it more reliable than existing single technical solutions.

[0052] The experimental verification of the present application on a large public data set shows that, compared with the average absolute error (Mean Absolute Error, MAE) of about 3.5 years of the existing mainstream scheme, the method of the present application can successfully reduce the MAE to 2.9 years; this precision improvement is due to the present application's ability to generate more comprehensive and robust brain feature representations; by fusing local details and global context, the model can capture complex structural patterns related to age that are ignored by single architecture models, thereby making a more physiological true age judgment, providing a more reliable biomarker for clinical diagnosis and health assessment, significantly improving the prediction accuracy and making the results more reliable.

[0053] The double-branch parallel structure and the core cross-attention gated fusion module designed by the present application overcome the problem of "partiality" in the prior art; it does not simply splice two features, but realizes intelligent interaction of information through cross-attention mechanism - let the global information guide the model to focus on key local details, and let the local details provide concrete evidence for abstract global relationships; the gating mechanism further realizes dynamic and adaptive fusion; this 1+1>2 synergistic effect makes the finally generated fusion features have stronger representation ability and discriminability than any single source feature, making the feature representation stronger and the information utilization more comprehensive.

[0054] The application fully considers the practicability of the model while pursuing high precision; the 3D-CNN branch adopts lightweight design, and the parameter quantity is only half of that of a standard network, effectively reducing the overall calculation burden and storage demand of the model; this makes the model of the application have a better performance-power ratio compared with some heavy networks (such as deep 3D-ResNet) that also pursue high performance but have bloated structures, is easier to train and deploy on a regular computing workstation, enhances the clinical conversion potential of the technical solution, and makes the calculation efficiency and model performance achieve a better balance.

[0055] The application adopts the guiding role of cross attention, which makes the global context features as a "compass" to guide the 3D-CNN features to focus on the local areas most related to the global aging pattern, avoiding the local branch from being interfered by irrelevant textures; on the contrary, the local detail features also provide specific anatomical basis for abstract global dependency, making the learning process of the Transformer more interpretable.

[0056] The introduction of the gating signal in the application enables the model to dynamically adjust the contribution proportion of local details and global context in the final decision according to the unique characteristics of each input sample; this data-driven adaptive ability is not possessed by simple fusion methods (such as splicing and addition), greatly improving the discriminant robustness of the features.

[0057] The above-described embodiments only express several embodiments of the application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be noted that for ordinary skilled persons in the art, without departing from the concept of the application, some modifications and improvements can be made, which are all within the protection scope of the application. Therefore, the protection scope of the patent of the application should be subject to the appended claims.

Claims

1. A brain age prediction method, characterized by, The method comprises the following steps: obtaining nuclear magnetic resonance image sMRI data of a target measurer; extracting local detail features representing the texture boundary of the cerebral cortex edge and representing the morphological features of the brain tissue in the nuclear magnetic resonance image sMRI data; slicing the nuclear magnetic resonance image sMRI data into image sequences, and extracting features in different feature channels for each image sequence, obtaining the dependency relationship between different image local regions based on windowed multi-head self-attention mechanism during feature extraction, and generating importance weights of different feature channels during global spatial feature extraction based on channel self-attention mechanism; according to the dependency relationship and the importance weights of different feature channels, the features extracted in different feature channels are weighted and fused to obtain global detail features; based on the cross-attention mechanism, the global detail features are used as keys for attention calculation, the global attention is obtained and weighted and fused to obtain local feature-rich global detail features, and the local detail features are used as queries for attention calculation, the local attention is obtained and weighted and fused to obtain local feature-enhanced local detail features; the original global detail features and the local detail features are spliced to obtain spliced features, and the spliced features are subjected to a full connection layer and a Sigmoid activation function to obtain dynamic gating weights; based on the dynamic gating weights, the local feature-rich global detail features and the local feature-enhanced local detail features are dynamically weighted and fused to obtain fusion features, and the brain age of the target measurer is predicted based on the fusion features.

2. The brain age prediction method of claim 1, wherein, The fusion features are obtained through a brain age prediction model, and the brain age prediction model comprises a cascaded lightweight 3D-CNN module, a Swin Transformer module and a cross-attention gating fusion module; The lightweight 3D-CNN module is composed of 6 basic units connected in series, each basic unit comprising a three-dimensional convolution layer with a convolution kernel size of 3x3x3 and a maximum pooling layer with a pooling kernel size of 2x2x2, for extracting local detail features; The Swin Transformer module is based on the Swin Transformer architecture, and the channel self-attention mechanism is introduced in parallel in each windowed multi-head self-attention block, for extracting global detail features; The cross-attention gating fusion module comprises a cascaded linear layer, a cross-attention mechanism and a full connection layer, wherein the activation function adopts a Sigmoid function, for extracting the local feature-rich global detail features and the local feature-enhanced local detail features, and dynamically weighting and fusing the local feature-rich global detail features and the local feature-enhanced local detail features.

3. The brain age prediction method of claim 2, wherein, The acquisition of the local detail features comprises: Feature extraction is performed by a three-dimensional convolution layer with a kernel size of 3x3x3 to obtain a three-dimensional feature map, and the three-dimensional feature map is aggregated into a one-dimensional deep convolution feature vector by a maximum pooling layer with a kernel size of 2x2x2 ; wherein, for the three-dimensional convolutional layer first layer, the value of any voxel on the output feature map of the three-dimensional convolution operation is represented as: ​ ； wherein: represents an input feature map; and represents the convolution kernel weight and bias of the layer; represents an activation function.

4. The brain age prediction method of claim 3, wherein, The acquisition of the global detail features comprises: For an input image sequence where is the sequence length, is the feature dimension, the self-attention computation is represented as: ； wherein: respectively represent the query, key and value matrices obtained by linear transformation; The windowed multi-head self-attention is limited in non-overlapping local windows, and the information interaction across windows is realized through window shift SW-MSA; On the basis of information interaction, a channel attention weight is introduced Element-wise multiplication is performed with the output features to enhance the feature representation, and a one-dimensional deep long-distance feature vector is output .

5. The brain age prediction method of claim 4, wherein, The acquisition of the fusion features comprises: by two independent linear layers and will and be projected to the same feature dimension, resulting in aligned features and ; By using cross attention mechanism, let the features and mutually rescale each other, use global information to guide local information , generate enhanced local features : ； with local information to enrich global information , generating enhanced global features is: ； The original alignment features and are stitched, and a gating network composed of two fully connected layers is input , and a gating signal is generated through a Sigmoid function , and the gating signal is represented as: ； Utilizing a gating signal The two features after re-calibration are dynamically weighted and summed to obtain the final fusion feature is: ； wherein: represents an element-wise multiplication.

6. The brain age prediction method of claim 1, wherein, After the nuclear magnetic resonance image sMRI data is obtained, the nuclear magnetic resonance image sMRI data is preprocessed, comprising: removing non-brain tissue in the nuclear magnetic resonance image sMRI to focus on the brain structure; After removing non-brain tissues, individual brains are aligned to standard space to realize the unification of voxel-level spatial coordinates. After spatial registration, field correction is performed to eliminate image intensity non-uniformity.

7. A brain age prediction system characterized by comprising: The method comprises the following steps: An image module is configured to acquire sMRI data of a target subject; A feature extraction module is configured to extract local detail features representing the texture boundary of the cerebral cortex edge and the morphological features of the brain tissue in the sMRI data; The sMRI data is sliced into image sequences, and feature extraction is performed on each image sequence under different feature channels. The importance weights of different feature channels are generated based on the windowed multi-head self-attention mechanism during global spatial feature extraction. The features extracted from different feature channels are weighted and fused according to the dependency relationship and the importance weights of different feature channels to obtain global detail features.

8. An electronic device, comprising: A cross-fusion module is configured to perform attention calculation based on the cross-attention mechanism, obtain global attention, and weightedly fuse to obtain local feature-rich global detail features. The original global detail features and local detail features are spliced to obtain spliced features, which are input into a full connection layer and a Sigmoid activation function to obtain dynamic gating weights. The local feature-rich global detail features and the local feature-enhanced local detail features are dynamically weighted and fused based on the dynamic gating weights to obtain fusion features, and the brain age of the target subject is predicted based on the fusion features. The method comprises the following steps:

9. A computer-readable storage medium, characterized in that, A memory and a processor are provided. The memory is configured to store a computer program. The processor is configured to execute the computer program stored in the memory to realize the steps of the brain age prediction method according to any one of claims 1-6. A computer program is stored in the memory, and the computer program is executed by the processor to realize the steps of the brain age prediction method according to any one of claims 1-6.

Images