Fine-grained brain age prediction method and system, terminal and storage medium

Abstract

The invention relates to the technical field of brain image analysis, and discloses a fine-grained brain age prediction method and system, a terminal and a storage medium, and the method comprises the steps: carrying out the preprocessing of a brain structure nuclear magnetic resonance image of a target object, and segmenting a plurality of brain tissue structures, according to the brain tissue structures, multiple feature maps are generated and mapped to multiple frontal bodies, mapped samples are processed through a four-stage network model, and dynamic self-adaptive lateral attention is added after each stage so as to predict the brain age of each brain region. According to the method, the brain age is predicted by utilizing the fine granularity of the cerebral cortex region level, and meanwhile, a dynamic self-adaptive lateral attention mechanism is introduced to simulate a real brain structure lateral relationship, so that the topological structure and local features of the cortex can be effectively expressed, and errors caused by brain form differences are reduced.

Description

Technical Field

[0001] This invention relates to the field of brain image analysis technology, and in particular to a fine-grained brain age prediction method, system, terminal, and storage medium. Background Technology

[0002] The development of the human brain is complex and dynamic. "Brain age" is a biological indicator that reflects the maturity of an individual's brain structure and function. It has been widely used in fields such as neurodevelopment and disease screening. It helps to accurately reflect the true state of an individual's neurodevelopment, assess the current level of brain development and damage, and facilitate early intervention.

[0003] However, existing brain age prediction methods do not take into account the physiological structure and characteristics of the brain, and often learn "image signals" rather than "neurodevelopmental information". This not only weakens the scientific validity and explanatory power of the model, but also limits its generalization and practical application in clinical scenarios.

[0004] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0005] The main objective of this invention is to provide a fine-grained method, system, terminal, and storage medium for predicting brain age, aiming to solve the problem that existing technologies do not consider the physiological structure and characteristics of the brain when predicting brain age, resulting in inaccurate brain age prediction results.

[0006] To achieve the above objectives, the present invention provides a fine-grained brain age prediction method, which includes the following steps: The initial MRI images of the target subject's brain are acquired, and the initial MRI images are segmented to obtain tissue types in different regions. Based on all the tissue types, curvature feature maps, cortical thickness feature maps, and sulcus depth feature maps are calculated. Based on the curvature feature map, the cortical thickness feature map, and the sulcus depth feature map, the brain is resampled onto a polyhedron to obtain multiple feature representations, and all the feature representations are converted into corresponding embedded features; All the embedded features are input into the constructed cross-hemispherical attention model, which converts all the embedded features into multiple global representations and concatenates all the global representations with the multiple embedded features to obtain multiple interactive features. After uniformly dividing all the interaction features, they are enhanced separately to obtain the enhanced features corresponding to each interaction feature. All the enhanced features are converted into multiple matrices. Based on all the matrices, multiple training features are generated through cross-attention mechanism and gating filtering mechanism. The training features are processed through multiple rounds of iterative processing using the cross-hemispheric attention model, and all feature representations are input into a multilayer perceptron for prediction, outputting the predicted brain age of the target object in different regions.

[0007] Optionally, in the fine-grained brain age prediction method, the tissue types include: gray matter, cerebrospinal fluid, white matter, and pia mater; The process involves acquiring initial MRI images of the target subject's brain, segmenting these images to obtain tissue types for different regions, and calculating curvature feature maps, cortical thickness feature maps, and sulcus depth feature maps based on all tissue types. Specifically, this includes: Acquire initial MRI images of the target subject's brain, perform head motion correction and registration on the initial MRI images, and remove non-brain tissue from the initial MRI images; The processed initial MRI image is segmented to obtain the gray matter, the cerebrospinal fluid, the white matter, and the pia mater; The boundary between the gray matter and the cerebrospinal fluid is extended to obtain a white matter mesh to generate a pia mater surface, and the sulcus curvature is estimated based on the surface of the white matter. The thickness of the gray skin layer is obtained by calculating the Euclidean distance between the groove curvature and the surface of the soft membrane. The surface of the white matter is expanded to generate an expanded surface, and the trench depth is estimated during the expansion. Based on the groove curvature, the gray cortex thickness, and the groove depth, a curvature feature map, a cortex thickness feature map, and a groove depth feature map are generated.

[0008] Optionally, in the fine-grained brain age prediction method, the feature representation includes: left brain feature representation and right brain feature representation; The process involves resampling the brain onto a polyhedron based on the curvature feature map, the cortical thickness feature map, and the sulcus depth feature map to obtain multiple feature representations, and then converting all of these feature representations into corresponding embedding features. Specifically, this includes: Based on the curvature feature map, the cortical thickness feature map, and the sulcus depth feature map, the brain is resampled onto a sixth-order regular icosahedron, and each face is divided into multiple triangular meshes to obtain the left brain feature representation and the right brain feature representation corresponding to each triangular mesh: ; in, The expression representing the feature. and These represent the characteristics of the left brain and the characteristics of the right brain, respectively. This indicates the number of triangular grids in each hemisphere. This indicates the number of vertices in each triangular mesh. The number of channels representing the feature. The dimension is The real number space; Each left-brain feature representation and each right-brain feature representation are expanded and input into a linear layer to obtain the corresponding embedded features.

[0009] Optionally, the fine-grained brain age prediction method, wherein the step of expanding each left-brain feature representation and each right-brain feature representation and inputting them into a linear layer to obtain the corresponding embedded features specifically includes: Expanding each left-brain feature representation and each right-brain feature representation yields the flattened left-brain and right-brain feature representations: ; in, The expression representing the characteristics of the left and right hemispheres after flattening. This indicates the characteristics of the flattened left brain. This indicates the characteristics of the right brain after it has been flattened. Representing dimensions The real space, Indicates will and Flattened into a one-dimensional plane; Each flattened left-brain feature representation and right-brain feature representation is mapped to multi-dimensional embedded features through a linear layer: ; in, Represents a sequence of all embedded features. and These represent the embedding features of the first triangular image patch in the left and right hemispheres, respectively. and Representing the first and second hemispheres respectively Embedding features of triangular image patches, Represents the weight matrix. This represents the dimension mapped through a linear layer. Representing dimensions The real number space.

[0010] Optionally, in the fine-grained brain age prediction method, the global representation includes: a left hemisphere global representation and a right hemisphere global representation; the interaction features include: left hemisphere interaction features and right hemisphere interaction features. The process involves inputting all the embedded features into a pre-constructed cross-hemispherical attention model. The cross-hemispherical attention model converts all the embedded features into multiple global representations, and then concatenates each of the global representations with one of the embedded features to obtain multiple interactive features, specifically including: A cross-hemispheric attention model is constructed. The sequence is input into the spatial mixer of the cross-hemispheric attention model. The spatial mixer performs random pooling operations on the left-brain embedding features and right-brain embedding features in the sequence respectively to obtain the left-hemispheric global representation and the right-hemispheric global representation. ; ; in, and These represent the global representations of the left and right hemispheres, respectively. This indicates a random pooling operation. , , and These represent different linear layers. This represents the first activation function. This represents all possible values ​​for the current dimension. This means taking the first half of the values ​​starting from the beginning of the current dimension. This indicates taking the second half of the value starting from the middle of the current dimension; The left hemisphere global representation is concatenated with all embedded features of the left hemisphere and then transformed to obtain the left hemisphere interaction features, which are then output. Similarly, the right hemisphere global representation is concatenated with all embedded features of the right hemisphere and then transformed to obtain the right hemisphere interaction features, which are then output. ; ; in, and These represent the interaction features of the left hemisphere and the interaction features of the right hemisphere, respectively. Indicates transpose. , , and These represent different linear layers. This indicates splicing / merging.

[0011] Optionally, in the fine-grained brain age prediction method, the enhancement features include: left hemisphere enhancement features and right hemisphere enhancement features; the training features include: left hemisphere training features and right hemisphere training features. The process involves uniformly dividing all the interaction features and then enhancing them separately to obtain enhanced features corresponding to each interaction feature. All enhanced features are then converted into multiple matrices. Based on all the matrices, multiple training features are generated using a cross-attention mechanism and a gating filtering mechanism. Specifically, this includes: The left hemisphere interaction features and the right hemisphere interaction features are input into the channel mixer of the cross-hemispheric attention model to uniformly segment the left hemisphere interaction features and the right hemisphere interaction features, resulting in multiple segmentation features: ; ; in, , and These represent the segmentation features of a uniformly divided left hemisphere. , and These represent the segmentation features of a uniformly divided right hemisphere. This indicates an operation that divides the data evenly. Indicates to Perform a uniform division operation. Indicates to Perform a uniform division operation; The first segmentation feature of the left and right hemispheres is processed by one-dimensional depthwise convolution, and all other segmentation features are processed by max pooling, one-dimensional depthwise convolution, and upsampling, respectively, to output the enhanced features of the left and right hemispheres. ; ; in, and These represent enhancement features in the left and right hemispheres, respectively. This represents the first activation function. This represents one-dimensional depthwise convolution processing. Indicates from the first The first left hemisphere segmentation feature or the second right hemisphere segmentation feature is spliced ​​together to the first... A left hemisphere segmentation feature or a right hemisphere segmentation feature. Indicates an upsampling operation. This indicates that independent convolution operations are performed on multiple channels. This indicates depthwise convolution processing. Indicates the first Max pooling is performed on the left hemisphere segmentation features. Indicates the first Max pooling is performed on the right hemisphere segmentation features. It represents the Hadamah accumulation. Indicates the number of segmentation features in the left hemisphere or the right hemisphere; The left hemisphere enhancement features and the right hemisphere enhancement features are respectively converted into the corresponding embedding matrix, query matrix, and reference matrix: , ; , ; , ; in, and These represent the conversion of the left hemisphere enhancement features and the right hemisphere enhancement features into their corresponding embedding matrices, respectively. and These represent the weight matrices corresponding to the embedding matrices of the left and right hemispheres, respectively. and These represent the conversion of the left hemisphere enhanced features and the right hemisphere enhanced features into corresponding query matrices, respectively. and These represent the weight matrices corresponding to the query matrices in the left and right hemispheres, respectively. and These represent the transformations of the left hemisphere enhancement features and the right hemisphere enhancement features into their corresponding reference matrices, respectively. and These represent the weight matrices corresponding to the reference matrices for the left and right hemispheres, respectively. Calculate the cross-attention mechanism based on all the reference matrices and all the query matrices: ; ; in, and These represent the cross-attention mechanisms targeting the left and right hemispheres, respectively. Represents the normalized exponential function, Indicates the dimensions of the query matrix and parameter matrix. For transpose, Indicates to Transpose. Indicates to Transpose; The embedding matrix and cross-attention mechanism corresponding to the left hemisphere are input into the first gating filter mechanism, and the embedding matrix and cross-attention mechanism corresponding to the right hemisphere are input into the second gating filter mechanism, which outputs the training features of the left hemisphere and the training features of the right hemisphere, respectively. ; ; in, and These represent the results after the first linear layer. and , Represents the hyperbolic tangent function. This represents the second activation function. and These represent the results after the second linear layer. and , and These represent the gating scores calculated for the left and right hemispheres, respectively. This represents the third activation function. and These represent the training characteristics of the left hemisphere and the right hemisphere, respectively. , , , , , , , , and These represent different linear layers.

[0012] Optionally, the fine-grained brain age prediction method, wherein the step of using the cross-hemispheric attention model to perform multiple rounds of iterative processing on all the training features, and inputting all the feature representations into a multilayer perceptron for prediction, outputting the predicted brain age of the target object in different regions, specifically includes: The training features of the left hemisphere and the training features of the right hemisphere are input into the next spatial mixer and channel mixer in the cross-hemispherical attention model for iterative processing until all processes are completed, and the final prediction data of each triangular grid is obtained. All the final prediction data are input into a multilayer perceptron for prediction, and the predicted brain age of each triangular grid of the target object is output: ; in, Indicates the first Predicted brain age using a triangular grid. Indicates a linear layer. Indicates the first The final predicted data.

[0013] Furthermore, to achieve the above objectives, the present invention also provides a fine-grained brain age prediction system, wherein the fine-grained brain age prediction system includes: The image preprocessing module is used to acquire the initial MRI image of the target object's brain, segment the initial MRI image to obtain multiple tissue types, and calculate curvature feature map, cortical thickness feature map and groove depth feature map according to all the tissue types. The spatial mixing module is used to resample the brain onto a polyhedron based on the curvature feature map, the cortical thickness feature map, and the sulcus depth feature map to obtain multiple feature representations, and to convert all the feature representations into corresponding embedded features; The channel mixing module is used to input all the embedded features into the constructed cross-hemispherical attention model. The cross-hemispherical attention model converts all the embedded features into multiple global representations and concatenates all the global representations with the multiple embedded features to obtain multiple interactive features. The dynamic adaptive bias module is used to uniformly divide all the interaction features and then perform enhancement processing on each of them to obtain the enhanced features corresponding to each interaction feature. All the enhanced features are converted into multiple matrices, and multiple training features are generated based on all the matrices through a cross-attention mechanism and a gating filtering mechanism. The brain age prediction module is used to perform multiple rounds of iterative processing on all the training features using the cross-hemispheric attention model, and input all the feature representations into a multilayer perceptron for prediction, and output the predicted brain age of the target object in different regions.

[0014] Furthermore, to achieve the above objectives, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and a fine-grained brain age prediction program stored in the memory and executable on the processor, wherein when the fine-grained brain age prediction program is executed by the processor, it implements the steps of the fine-grained brain age prediction method as described above.

[0015] Furthermore, to achieve the above objectives, the present invention also provides a storage medium storing a fine-grained brain age prediction program, which, when executed by a processor, implements the steps of the fine-grained brain age prediction method as described above.

[0016] In this invention, initial MRI images of the target object's brain are acquired, and the initial MRI images are segmented to obtain tissue types in different regions. Based on all tissue types, curvature feature maps, cortical thickness feature maps, and sulcus depth feature maps are calculated. Based on the curvature feature maps, cortical thickness feature maps, and sulcus depth feature maps, the brain is resampled onto a polyhedron to obtain multiple feature representations, and all feature representations are converted into corresponding embedded features. All embedded features are input into a pre-constructed cross-hemispheric attention model, which converts all embedded features into multiple global representations, and concatenates all global representations with the multiple embedded features to obtain multiple interactive features. All interactive features are uniformly divided and enhanced to obtain enhanced features corresponding to each interactive feature. All enhanced features are converted into multiple matrices, and based on all matrices, multiple training features are generated through a cross-attention mechanism and a gating filtering mechanism. The cross-hemispheric attention model is used to perform multiple rounds of iterative processing on all training features, and all feature representations are input into a multilayer perceptron for prediction, outputting the predicted brain age of the target object in different regions. This invention uses fine-grained regional-level analysis of the cerebral cortex to predict brain age, while introducing a dynamic adaptive lateral attention mechanism to simulate the lateral relationship of real brain structures. This effectively expresses the topological structure and local features of the cortex and reduces errors caused by differences in brain morphology. Attached Figure Description

[0017] Figure 1 This is a flowchart of a preferred embodiment of the fine-grained brain age prediction method of the present invention; Figure 2 This is a flowchart illustrating a preferred embodiment of the fine-grained brain age prediction method of the present invention. Figure 3 This is a schematic diagram illustrating the developmental level of a preterm population, representing a preferred embodiment of the fine-grained brain age prediction method of the present invention. Figure 4 This is a diagram illustrating the developmental differences between different regions of the left and right hemispheres in a preterm infant population, representing a preferred embodiment of the fine-grained brain age prediction method of the present invention. Figure 5 This is a whole-brain difference map of a preterm population, which is a preferred embodiment of the fine-grained brain age prediction method of the present invention. Figure 6 This is a schematic diagram illustrating the developmental level of extremely premature infants, representing a preferred embodiment of the fine-grained brain age prediction method of the present invention. Figure 7 This is a diagram showing the developmental differences between different regions of the left and right hemispheres in a preterm infant population, representing a preferred embodiment of the fine-grained brain age prediction method of the present invention. Figure 8This is a whole-brain difference map of an extremely premature infant population, which is a preferred embodiment of the fine-grained brain age prediction method of the present invention. Figure 9 This is a schematic diagram of the developmental level of a localized developmental abnormality population, representing a preferred embodiment of the fine-grained brain age prediction method of the present invention. Figure 10 This is a preferred embodiment of the fine-grained brain age prediction method of the present invention, showing the developmental differences between different regions of the left and right hemispheres in a population with local developmental abnormalities. Figure 11 This is a preferred embodiment of the fine-grained brain age prediction method of the present invention, showing the whole-brain differences in individuals with local developmental abnormalities. Figure 12 This is a comparison chart of high- and low-risk autism populations, representing a preferred embodiment of the fine-grained brain age prediction method of the present invention. Figure 13 This is a structural diagram of a preferred embodiment of the fine-grained brain age prediction system of the present invention; Figure 14 This is a structural diagram of a preferred embodiment of the terminal of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] The development of the human brain is complex and dynamic. "Brain age" is a biological indicator that reflects the maturity of an individual's brain structure and function. It has been widely used in fields such as neurodevelopment and disease screening. It helps to accurately reflect the true state of an individual's neurodevelopment, assess the current level of brain development and damage, and facilitate early intervention.

[0020] With the continuous development of deep learning technology, researchers have begun to explore combining traditional data processing methods with new algorithms to improve the ability to analyze complex data. Some of these works primarily use voxel-level features of imaging data such as SMRI (Structural Magnetic Resonance Imaging), DMRI (Diffusion Magnetic Resonance Imaging), and fMRI (Functional Magnetic Resonance Imaging) to directly predict brain age, achieving good results. However, these methods typically ignore the spatial topology of the cerebral cortex itself and cannot effectively model the geometric connections and developmental associations between cortical regions.

[0021] Mapping brain structural features onto a sphere allows for better capture of the brain's geometric features and connectivity patterns. However, existing technologies only predict the overall age of the brain without considering developmental deviations in fine-grained brain regions, making it difficult to reveal potential neurodevelopmental abnormalities reflected by developmental inconsistencies between regions. Therefore, this invention proposes a cross-hemispheric attention network modeling using structural lateralization and region modeling to model the developmental trajectories of different regions of the whole cerebral cortex. This model has a mean absolute error of 0.54 ± 0.04 (mean ± standard deviation) for predicting the global brain age of newborns and a mean absolute error of 0.45 ± 0.39 (mean ± standard deviation) for predicting fine-grained cortical regions, demonstrating high prediction accuracy.

[0022] The fine-grained brain age prediction method described in the preferred embodiment of the present invention, such as... Figure 1 As shown, the fine-grained brain age prediction method includes the following steps: Step S10: Obtain the initial MRI image of the target object's brain, segment the initial MRI image to obtain the tissue types of different regions, and calculate the curvature feature map, cortical thickness feature map and sulcus depth feature map based on all the tissue types.

[0023] In the embodiments disclosed in this invention, the data used is MRI images of the brain structure of a newborn (i.e., the target object). The reconstructed T2-weighted images after motion correction (referring to the imaging in MRI that displays tissue characteristics by highlighting differences in transverse tissue relaxation (T2 relaxation)) are biased and brain tissue is extracted. Subsequently, the brain images are segmented into different tissue types (cerebrospinal fluid, white matter, cortical gray matter, and subcutaneous gray matter). The white matter surface is expanded to fit the pia mater surface. A medium-thickness surface is fitted between the white matter and the pia mater and the thickness is estimated. The white matter surface is expanded to fit the expanded surface and projected onto a sphere. The curvature is estimated from the white matter surface. The brain sulcus depth map (mean convexity / concavity) is estimated.

[0024] When acquiring image data, it is necessary to ensure that all acquisition parameters remain consistent, then preprocess the data, and then use a preset algorithm to segment the brain image into multiple tissue types, such as gray matter, cerebrospinal fluid, white matter, and pia mater (including but not limited to these); then feature maps of various fine-grained levels of the target object can be extracted.

[0025] Specifically, the initial MRI images of the target subject's brain are acquired; head motion correction and registration are performed on the initial MRI images, and non-brain tissues are removed from the initial MRI images; the processed initial MRI images are segmented to obtain the gray matter, the cerebrospinal fluid, the white matter, and the pia mater; the boundaries of the gray matter and the cerebrospinal fluid are expanded to obtain a white matter mesh to generate the pia mater surface, and the sulcus curvature is estimated based on the surface of the white matter; the Euclidean distance between the sulcus curvature and the pia mater surface is calculated to obtain the gray matter cortical thickness; the surface of the white matter is expanded to generate an expanded surface, and the sulcus depth is estimated during the expansion; based on the sulcus curvature, the gray matter cortical thickness, and the sulcus depth, a curvature feature map, a cortical thickness feature map, and a sulcus depth feature map are generated.

[0026] The image preprocessing includes head motion correction, registration, non-brain tissue removal, gray matter boundary reconstruction, and deviation correction to remove non-detection data from the image, thereby improving the accuracy of subsequent brain age prediction based on various fine-grained feature maps. After preprocessing, the image is segmented to obtain different tissue types in different regions, including cerebrospinal fluid, white matter, cortical gray matter, and subcortical gray matter. These tissue types reflect the characteristics of different regions of the cerebral cortex. Then, various feature maps are extracted using these tissue types to predict the brain age of different regions of the cerebral cortex, achieving brain age prediction at a fine-grained level: the pia mater surface is generated by extending the white matter mesh to the gray matter and cerebrospinal fluid boundaries; the sulcus curvature is estimated from the white matter surface; the gray matter cortical thickness is estimated based on the Euclidean distance between the white matter surface and the pia mater surface; an expanded surface is generated by smoothing the white matter surface based on the expansion, and the sulcus depth is estimated during the expansion; finally, curvature feature maps, cortical thickness feature maps, and sulcus depth feature maps are generated based on the sulcus curvature, gray matter cortical thickness, and sulcus depth.

[0027] Step S20: Based on the curvature feature map, the cortical thickness feature map, and the sulcus depth feature map, the brain is resampled onto a polyhedron to obtain multiple feature representations, and all the feature representations are converted into corresponding embedded features.

[0028] The feature representations include left-brain and right-brain feature representations. After extracting these feature maps, the brain is resampled to divide it into multiple regions, which can more effectively represent the long-term co-developmental relationship between different brain regions and make up for the shortcomings of traditional CNN architecture (Convolutional Neural Network) in modeling spatial dependencies.

[0029] Specifically, based on the curvature feature map, the cortical thickness feature map, and the sulcus depth feature map, the brain is resampled onto a sixth-order regular icosahedron, and each face is divided into multiple triangular grids to obtain the left brain feature representation and the right brain feature representation corresponding to each triangular grid: ; in, The expression representing the feature. and These represent the characteristics of the left brain and the characteristics of the right brain, respectively. This indicates the number of triangular grids in each hemisphere. This indicates the number of vertices in each triangular mesh. The number of channels representing the feature. The dimension is The real number space; after expanding each left brain feature representation and each right brain feature representation, input them into a linear layer to obtain the corresponding embedded features.

[0030] In the embodiments disclosed in this invention, after extracting feature maps of sulcus depth, curvature, and cortical thickness, the surfaces of the left and right hemispheres are resampled onto a sixth-order regular icosahedron and divided into 5120 triangular grids, each grid containing 15 vertices. The feature channels include sulcus depth, curvature, and cortical thickness. By sampling the spherical surface of the brain into multiple triangular grids, the effect of individually predicting each region is achieved. Fine-grained regions of the cerebral cortex possess good physiological continuity and spatial consistency, and can more effectively express the topological structure and local features of the cortex.

[0031] Furthermore, each of the left-brain feature representations and each of the right-brain feature representations is unfolded to obtain the flattened left-brain feature representations and right-brain feature representations: ; in, The expression representing the characteristics of the left and right hemispheres after flattening. This indicates the characteristics of the flattened left brain. This indicates the characteristics of the right brain after it has been flattened. Representing dimensions The real space, Indicates will and Flattened into a one-dimensional plane; each flattened left-brain feature representation and right-brain feature representation is mapped to multi-dimensional embedded features through a linear layer: ; in, Represents a sequence of all embedded features. and These represent the embedding features of the first triangular image patch in the left and right hemispheres, respectively. and Representing the first and second hemispheres respectively Embedding features of triangular image patches, Represents the weight matrix. This represents the dimension mapped through a linear layer. Representing dimensions The real number space.

[0032] The U-Net architecture (Convolutional Neural Network for Biomedical Image Segmentation), commonly used in this field, suffers from limitations in its local receptive field due to the inherent limitations of CNNs, making it insufficient for modeling co-developmental relationships between distant brain regions. Although stacking more layers or expanding the receptive field can alleviate this problem to some extent, it is still difficult to effectively capture non-local structural dependencies, which often play a crucial role in neurodevelopmental disorders such as autism and attention deficit hyperactivity disorder.

[0033] Therefore, this invention proposes a cross-hemispheric attention network (i.e., the cross-hemispheric attention model below) with region modeling. The network consists of four stages, each of which contains a Ni-layer spatial channel mixing module (where i=1, 2, 3, 4; N1, N2, N3, N4=[2, 2, 6, 2]). By introducing a module with global modeling capabilities, it can more effectively represent the long-term co-developmental relationship between different brain regions, making up for the shortcomings of traditional CNN architecture in spatial dependency modeling capabilities.

[0034] For the input to the cross-hemispheric attention network, the left-brain feature representation and the right-brain feature representation obtained by resampling need to be expanded separately and then fed into a linear layer to be mapped into multi-dimensional embedding features. For the embedding features of each triangular grid (including left-brain and right-brain), they are paired to form a sequence, which is used to input the cross-hemispheric attention network.

[0035] Step S30: Input all the embedded features into the constructed cross-hemispherical attention model. The cross-hemispherical attention model converts all the embedded features into multiple global representations and concatenates all the global representations with the multiple embedded features to obtain multiple interactive features.

[0036] The global representation includes: left hemisphere global representation and right hemisphere global representation; the interaction features include: left hemisphere interaction features and right hemisphere interaction features; for the input embedded features, four stages of the cross-hemispheric attention network are required (e.g., Figure 2As shown in the figure, each stage is followed by a dynamic adaptive lateral attention mechanism.

[0037] Specifically, a cross-hemispheric attention model is constructed, and the sequence is input into the spatial mixer of the cross-hemispheric attention model. The spatial mixer performs random pooling operations on the left-brain embedding features and the right-brain embedding features in the sequence respectively to obtain the left-hemispheric global representation and the right-hemispheric global representation. ; ; in, and These represent the global representations of the left and right hemispheres, respectively. This indicates a random pooling operation. , , and These represent different linear layers. This represents the first activation function. This represents all possible values ​​for the current dimension. This means taking the first half of the values ​​starting from the beginning of the current dimension. This indicates taking the second half of the value starting from the middle of the current dimension; the left hemisphere global representation is concatenated with all embedded features of the left hemisphere and then transformed to obtain the left hemisphere interaction feature, which is then output; the right hemisphere global representation is concatenated with all embedded features of the right hemisphere and then transformed to obtain the right hemisphere interaction feature, which is then output. ; ; in, and These represent the interaction features of the left hemisphere and the interaction features of the right hemisphere, respectively. Indicates transpose. , , and These represent different linear layers. This indicates splicing / merging.

[0038] Each spatial-channel mixing module in each stage includes a spatial mixer and a channel mixer. A separate spatial-channel mixing module is used for each hemisphere; the spatial mixer models long-range dependencies within the hemisphere, and then the channel mixer enhances the representation of each region to predict the brain age of each region.

[0039] In this method, voxel-level prediction directly acts on the original image pixels without incorporating the physiological structural information of the cortex. This makes it susceptible to interference from factors such as scanning noise, image registration errors, non-brain tissues (such as background, skull, etc.), or structural signals unrelated to brain age prediction, thus affecting prediction performance. Therefore, in this embodiment of the present invention, fine-grained regions of the cerebral cortex are used for preprocessing of brain age prediction, avoiding the training difficulties and prediction errors brought about by using voxel-level models. This method can more effectively express the topological structure and local features of the cortex. Furthermore, these regions can achieve consistent alignment across individuals through standard cortical templates, thereby reducing errors caused by differences in brain morphology.

[0040] Step S40: After uniformly dividing all the interaction features, perform enhancement processing on each of them to obtain the enhanced features corresponding to each interaction feature. Convert all the enhanced features into multiple matrices. Based on all the matrices, generate multiple training features through cross-attention mechanism and gating filtering mechanism.

[0041] The enhancement features include left hemisphere enhancement features and right hemisphere enhancement features; the training features include left hemisphere training features and right hemisphere training features; in addition to the above-mentioned spatial channel mixing module, the cross-hemispheric attention network disclosed in this invention also includes cross-attention and gating filtering mechanisms (i.e., the above-mentioned dynamic adaptive lateral attention mechanism). The cross-attention mechanism mimics the selective attention of biological vision and can establish long-distance dependencies between hemispheres. The gating filtering mechanism filters contralateral attention and lateral information to simulate the laterality of brain structure.

[0042] Specifically, the left hemisphere interaction features and the right hemisphere interaction features are input into the channel mixer of the cross-hemispheric attention model to uniformly segment the left hemisphere interaction features and the right hemisphere interaction features, resulting in multiple segmentation features: ; ; in, , and These represent the segmentation features of a uniformly divided left hemisphere. , and These represent the segmentation features of a uniformly divided right hemisphere. This indicates an operation that divides the data evenly. Indicates to Perform a uniform division operation. Indicates to Perform uniform segmentation; apply one-dimensional depthwise convolution to the first segmentation feature of the left and right hemispheres, and perform max pooling, one-dimensional depthwise convolution, and upsampling on all other segmentation features respectively, outputting the enhanced features for the left and right hemispheres respectively: ; ; in, and These represent enhancement features in the left and right hemispheres, respectively. This represents the first activation function (Gaussian Error Linear Unit, an activation function based on a Gaussian distribution). This represents one-dimensional depthwise convolution processing. Indicates from the first The first left hemisphere segmentation feature or the second right hemisphere segmentation feature is spliced ​​together to the first... A left hemisphere segmentation feature or a right hemisphere segmentation feature. Indicates an upsampling operation. This indicates that independent convolution operations are performed on multiple channels. This indicates depthwise convolution processing. Indicates the first Max pooling is performed on the left hemisphere segmentation features. Indicates the first Max pooling is performed on the right hemisphere segmentation features. It represents the Hadamah accumulation. This indicates the number of left hemisphere segmentation features or right hemisphere segmentation features; the left hemisphere enhancement features and right hemisphere enhancement features are respectively converted into corresponding embedding matrices, query matrices, and reference matrices: , ; , ; , ; in, and These represent the conversion of the left hemisphere enhancement features and the right hemisphere enhancement features into their corresponding embedding matrices, respectively. and These represent the weight matrices corresponding to the embedding matrices of the left and right hemispheres, respectively. and These represent the conversion of the left hemisphere enhanced features and the right hemisphere enhanced features into corresponding query matrices, respectively. and These represent the weight matrices corresponding to the query matrices in the left and right hemispheres, respectively. and These represent the transformations of the left hemisphere enhancement features and the right hemisphere enhancement features into their corresponding reference matrices, respectively. and Let each represent a weight matrix corresponding to the reference matrix for the left and right hemispheres, respectively; and calculate the cross-attention mechanism based on all the reference matrices and all the query matrices. ; ; in, and These represent the cross-attention mechanisms targeting the left and right hemispheres, respectively. Represents the normalized exponential function, Indicates the dimensions of the query matrix and parameter matrix. For transpose, Indicates to Transpose. Indicates to Perform transposition; input the embedding matrix and cross-attention mechanism corresponding to the left hemisphere into the first gating filter mechanism, and input the embedding matrix and cross-attention mechanism corresponding to the right hemisphere into the second gating filter mechanism, and output the training features of the left hemisphere and the training features of the right hemisphere respectively: ; ; in, and These represent the results after the first linear layer. and , Represents the hyperbolic tangent function. This represents the second activation function (S-shaped function). and These represent the results after the second linear layer. and , and These represent the gating scores calculated for the left and right hemispheres, respectively. This represents the third activation function (Rectified Linear Unit, one of the most commonly used non-linear activation functions in deep neural networks). and These represent the training characteristics of the left hemisphere and the right hemisphere, respectively. , , , , , , , , and These represent different linear layers.

[0043] Among them, such as Figure 2 As shown, groove depth, curvature, and cortical thickness have been nonlinearly fused in the aforementioned nonlinear process. Therefore, in the channel mixer, the features of each region are first uniformly divided, and then the first feature ( and The process uses one-dimensional depthwise convolution with a kernel size of 3, and the remaining parts are processed by max pooling, one-dimensional depthwise convolution, and upsampling operations, respectively.

[0044] First, the data from the two hemispheres (i.e., the interaction features of the left hemisphere and the interaction features of the right hemisphere) are converted into three different matrices (e.g., ...). Figure 2 In this model, E, R, and Q represent the embedding matrix, reference matrix, and query matrix, respectively. Furthermore, the reference matrix also represents the key-value matrix and numerical matrix in the attention mechanism. Subsequently, a cross-attention mechanism is constructed based on the query matrix and reference matrix and classified according to the left and right hemispheres. At the same time, it is sent to the gated filtering mechanism along with the embedding matrix (i.e., the embedding matrix and cross-attention mechanism of the left hemisphere are sent to one gated filtering mechanism, and the embedding matrix and cross-attention mechanism of the right hemisphere are sent to another gated filtering mechanism). Finally, the input data for the next stage (i.e., the training features of the left hemisphere and the training features of the right hemisphere) are output.

[0045] Furthermore, in the embodiments disclosed in this invention, modeling is based on the cortical surface, which removes most of the interfering factors and provides finer-grained predictions, thereby improving the neuroscientific interpretability and application value of the prediction results. At the same time, using regions as modeling units not only improves the stability of the predictions but also makes the results structurally traceable and cognitively interpretable.

[0046] Step S50: Use the cross-hemispheric attention model to perform multiple rounds of iterative processing on all the training features, and input all the feature representations into a multilayer perceptron for prediction, and output the predicted brain age of the target object in different regions.

[0047] Specifically, the training features from the left and right hemispheres are input into the next spatial mixer and channel mixer in the cross-hemispheric attention model for iterative processing until all processes are completed, yielding the final prediction data for each triangular grid. All the final prediction data are then input into a multilayer perceptron for prediction, outputting the predicted brain age for each triangular grid of the target object. ; in, Indicates the first Predicted brain age using a triangular grid. Indicates a linear layer. Indicates the first The final predicted data.

[0048] After completing all stages of data processing, the data is processed through a multilayer perceptron (i.e., Figure 2 The MLP (Multilayer Perceptron) in the brain uses these data to make predictions, thereby obtaining the predicted brain age of the target object in each triangular grid region, and realizing the prediction of different regions of the whole cerebral cortex.

[0049] Furthermore, in another embodiment disclosed in this invention, such as Figure 3 , Figure 4 and Figure 5 As shown, the test results of the preterm infant population using the method of the present invention were presented. It was found that the left hemisphere of the preterm infant population showed a slight trend of delayed development, while a few areas of the right hemisphere showed a slight trend of overdevelopment.

[0050] Furthermore, in another embodiment disclosed in this invention, such as Figure 6 , Figure 7 and Figure 8 As shown, the test results of the method of the present invention on the developmental population of extremely premature infants were presented, and it was found that the whole brain of the extremely premature infant population showed obvious delayed development.

[0051] Furthermore, in another embodiment disclosed in this invention, such as Figure 9 , Figure 10 and Figure 11 As shown, the test results of a population with localized developmental abnormalities using the method of the present invention were demonstrated, revealing that the entire brain of the population with localized developmental abnormalities showed significant delayed development.

[0052] Furthermore, in another embodiment disclosed in this invention, such as Figure 12 As shown, a comparison was made between healthy subjects (low risk) and those at higher risk of autism. The results showed that, on the whole brain, the difference between the predicted brain age and the actual brain age was significantly higher in the high-risk group than in the low-risk group.

[0053] This invention uses fine-grained regional-level analysis of the cerebral cortex to predict brain age, while introducing a dynamic adaptive lateral attention mechanism to simulate the lateral relationship of real brain structures. This effectively expresses the topological structure and local features of the cortex and reduces errors caused by differences in brain morphology.

[0054] Furthermore, such as Figure 13 As shown, based on the above-described fine-grained brain age prediction method, the present invention also provides a fine-grained brain age prediction system, wherein the fine-grained brain age prediction system includes: Image preprocessing module 51 is used to acquire the initial MRI image of the brain of the target object, segment the initial MRI image to obtain the tissue types of different regions, and calculate the curvature feature map, cortical thickness feature map and groove depth feature map according to all the tissue types. The spatial mixing module 52 is used to resample the brain onto a polyhedron based on the curvature feature map, the cortical thickness feature map, and the groove depth feature map to obtain multiple feature representations, and to convert all the feature representations into corresponding embedded features; The channel mixing module 53 is used to input all the embedded features into the constructed cross-hemispherical attention model. The cross-hemispherical attention model converts all the embedded features into multiple global representations and concatenates all the global representations with the multiple embedded features to obtain multiple interactive features. The dynamic adaptive bias module 54 is used to uniformly divide all the interaction features and perform enhancement processing on each of them to obtain the enhanced features corresponding to each interaction feature, convert all the enhanced features into multiple matrices, and generate multiple training features based on all the matrices through a cross-attention mechanism and a gating filtering mechanism. The brain age prediction module 55 is used to perform multiple rounds of iterative processing on all the training features using the cross-hemispheric attention model, and input all the feature representations into a multilayer perceptron for prediction, and output the predicted brain age of the target object in different regions.

[0055] Furthermore, such as Figure 14 As shown, based on the above-mentioned fine-grained brain age prediction method and system, the present invention also provides a terminal, which includes a processor 10, a memory 20 and a display 30. Figure 14 Only some of the terminal components are shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0056] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc. Further, the memory 20 may include both internal and external storage devices. The memory 20 is used to store application software and various types of data installed on the terminal, such as program code installed on the terminal. The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores a fine-grained brain age prediction program 40, which can be executed by the processor 10 to implement the fine-grained brain age prediction method of this application.

[0057] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the fine-grained brain age prediction method.

[0058] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface. The components of the terminal communicate with each other via a system bus.

[0059] In one embodiment, the processor 10 implements the steps of the fine-grained brain age prediction method as described above when executing the fine-grained brain age prediction program 40 in the memory 20.

[0060] The present invention also provides a storage medium (i.e., a computer-readable storage medium) wherein the storage medium stores a fine-grained brain age prediction program, which, when executed by a processor, implements the steps of the fine-grained brain age prediction method as described above.

[0061] In summary, this invention provides a fine-grained brain age prediction method, system, terminal, and storage medium. The method includes: acquiring initial MRI images of the brain of a target object; segmenting the initial MRI images to obtain tissue types in different regions; calculating curvature feature maps, cortical thickness feature maps, and sulcus depth feature maps based on all tissue types; resampling the brain onto a polyhedron based on the curvature feature maps, cortical thickness feature maps, and sulcus depth feature maps to obtain multiple feature representations; converting all feature representations into corresponding embedded features; and inputting all embedded features into a pre-constructed cross-hemispheric attention model. The hemispherical attention model converts all embedded features into multiple global representations, and concatenates each global representation with one of the embedded features to obtain multiple interactive features. These interactive features are then uniformly divided and enhanced to obtain enhanced features for each interactive feature. All enhanced features are converted into multiple matrices, and based on these matrices, multiple training features are generated using a cross-attention mechanism and a gating filtering mechanism. The hemispherical attention model is used to iteratively process all training features in multiple rounds, and all feature representations are input into a multilayer perceptron for prediction, outputting the predicted brain age of the target object in different regions. This invention utilizes fine-grained regional-level brain age prediction and introduces a dynamic adaptive lateral attention mechanism to simulate the lateral relationships of real brain structures. This effectively expresses the topological structure and local features of the cortex, reducing errors caused by differences in brain morphology.

[0062] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal that includes that element.

[0063] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The storage medium can be a memory, magnetic disk, optical disk, etc.

[0064] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A fine-grained method for predicting brain age, characterized in that, The fine-grained brain age prediction method includes: The initial MRI images of the target subject's brain are acquired, and the initial MRI images are segmented to obtain tissue types in different regions. Based on all the tissue types, curvature feature maps, cortical thickness feature maps, and sulcus depth feature maps are calculated. Based on the curvature feature map, the cortical thickness feature map, and the sulcus depth feature map, the brain is resampled onto a polyhedron to obtain multiple feature representations, and all the feature representations are converted into corresponding embedded features; All the embedded features are input into the constructed cross-hemispherical attention model, which converts all the embedded features into multiple global representations and concatenates all the global representations with the multiple embedded features to obtain multiple interactive features. After uniformly dividing all the interaction features, they are enhanced separately to obtain the enhanced features corresponding to each interaction feature. All the enhanced features are converted into multiple matrices. Based on all the matrices, multiple training features are generated through cross-attention mechanism and gating filtering mechanism. The training features are processed through multiple rounds of iterative processing using the cross-hemispheric attention model, and all feature representations are input into a multilayer perceptron for prediction, outputting the predicted brain age of the target object in different regions.

2. The fine-grained brain age prediction method according to claim 1, characterized in that, The tissue types include: gray matter, cerebrospinal fluid, white matter, and pia mater; The process involves acquiring initial MRI images of the target subject's brain, segmenting these images to obtain tissue types for different regions, and calculating curvature feature maps, cortical thickness feature maps, and sulcus depth feature maps based on all tissue types. Specifically, this includes: Acquire initial MRI images of the target subject's brain, perform head motion correction and registration on the initial MRI images, and remove non-brain tissue from the initial MRI images; The processed initial MRI image is segmented to obtain the gray matter, the cerebrospinal fluid, the white matter, and the pia mater; The boundary between the gray matter and the cerebrospinal fluid is extended to obtain a white matter mesh to generate a pia mater surface, and the sulcus curvature is estimated based on the surface of the white matter. The thickness of the gray skin layer is obtained by calculating the Euclidean distance between the groove curvature and the surface of the soft membrane. The surface of the white matter is expanded to generate an expanded surface, and the trench depth is estimated during the expansion. Based on the groove curvature, the gray cortex thickness, and the groove depth, a curvature feature map, a cortex thickness feature map, and a groove depth feature map are generated.

3. The fine-grained brain age prediction method according to claim 1, characterized in that, The feature representation includes: left brain feature representation and right brain feature representation; The process involves resampling the brain onto a polyhedron based on the curvature feature map, the cortical thickness feature map, and the sulcus depth feature map to obtain multiple feature representations, and then converting all of these feature representations into corresponding embedding features. Specifically, this includes: Based on the curvature feature map, the cortical thickness feature map, and the sulcus depth feature map, the brain is resampled onto a sixth-order regular icosahedron, and each face is divided into multiple triangular meshes to obtain the left brain feature representation and the right brain feature representation corresponding to each triangular mesh: ； in, The expression representing the feature. and These represent the characteristics of the left brain and the characteristics of the right brain, respectively. This indicates the number of triangular grids in each hemisphere. This indicates the number of vertices in each triangular mesh. The number of channels representing the feature. The dimension is The real number space; Each left-brain feature representation and each right-brain feature representation are expanded and input into a linear layer to obtain the corresponding embedded features.

4. The fine-grained brain age prediction method according to claim 3, characterized in that, The step of expanding each left-brain feature representation and each right-brain feature representation and inputting them into a linear layer to obtain the corresponding embedding features specifically includes: Each left-brain feature representation and each right-brain feature representation are unfolded to obtain the flattened left-brain feature representation and right-brain feature representation: ； in, The expression representing the characteristics of the left and right hemispheres after flattening. This indicates the characteristics of the flattened left brain. This indicates the characteristics of the right brain after it has been flattened. Representing dimensions The real space, Indicates will and Flattened into a one-dimensional plane; Each flattened left-brain feature representation and right-brain feature representation is mapped to multi-dimensional embedded features through a linear layer: ； in, Represents a sequence of all embedded features. and These represent the embedding features of the first triangular image patch in the left and right hemispheres, respectively. and Representing the first and second hemispheres respectively Embedding features of triangular image patches, Represents the weight matrix. This represents the dimension mapped through a linear layer. Representing dimensions The real number space.

5. The fine-grained brain age prediction method according to claim 4, characterized in that, The global representation includes: a left hemisphere global representation and a right hemisphere global representation; the interactive features include: left hemisphere interactive features and right hemisphere interactive features; The process involves inputting all the embedded features into a cross-hemispherical attention model, which converts all the embedded features into multiple global representations. These global representations are then concatenated with the multiple embedded features to obtain multiple interactive features, specifically including: A cross-hemispheric attention model is constructed. The sequence is input into the spatial mixer of the cross-hemispheric attention model. The spatial mixer performs random pooling operations on the left-brain embedding features and right-brain embedding features in the sequence respectively to obtain the left-hemispheric global representation and the right-hemispheric global representation. ； ； in, and These represent the global representations of the left and right hemispheres, respectively. This indicates a random pooling operation. , , and These represent different linear layers. This represents the first activation function. This represents all possible values ​​for the current dimension. This means taking the first half of the values ​​starting from the beginning of the current dimension. This indicates taking the second half of the value starting from the middle of the current dimension; The left hemisphere global representation is concatenated with all embedded features of the left hemisphere and then transformed to obtain the left hemisphere interaction features, which are then output. Similarly, the right hemisphere global representation is concatenated with all embedded features of the right hemisphere and then transformed to obtain the right hemisphere interaction features, which are then output. ； ； in, and These represent the interaction features of the left hemisphere and the interaction features of the right hemisphere, respectively. Indicates transpose. , , and These represent different linear layers. This indicates splicing / merging.

6. The fine-grained brain age prediction method according to claim 5, characterized in that, The enhancement features include: left hemisphere enhancement features and right hemisphere enhancement features; the training features include: left hemisphere training features and right hemisphere training features; The process involves uniformly dividing all the interaction features and then enhancing them separately to obtain enhanced features corresponding to each interaction feature. All enhanced features are then converted into multiple matrices. Based on all the matrices, multiple training features are generated using a cross-attention mechanism and a gating filtering mechanism. Specifically, this includes: The left hemisphere interaction features and the right hemisphere interaction features are input into the channel mixer of the cross-hemispheric attention model to uniformly segment the left hemisphere interaction features and the right hemisphere interaction features, resulting in multiple segmentation features: ； ； in, , and These represent the segmentation features of a uniformly divided left hemisphere. , and These represent the segmentation features of a uniformly divided right hemisphere. This indicates an operation that divides the data evenly. Indicates to Perform a uniform division operation. Indicates to Perform a uniform division operation; The first segmentation feature of the left and right hemispheres is processed by one-dimensional depthwise convolution, and all other segmentation features are processed by max pooling, one-dimensional depthwise convolution, and upsampling, respectively, to output the enhanced features of the left and right hemispheres. ； ； in, and These represent enhancement features in the left and right hemispheres, respectively. This represents one-dimensional depthwise convolution processing. Indicates from the first The first left hemisphere segmentation feature or the second right hemisphere segmentation feature is spliced ​​together to the first... A left hemisphere segmentation feature or a right hemisphere segmentation feature. Indicates an upsampling operation. This indicates that independent convolution operations are performed on multiple channels. This indicates depthwise convolution processing. Indicates the first Max pooling is performed on the left hemisphere segmentation features. Indicates the first Max pooling is performed on the right hemisphere segmentation features. It represents the Hadamah accumulation. Indicates the number of segmentation features in the left hemisphere or the right hemisphere; The left hemisphere enhancement features and the right hemisphere enhancement features are respectively converted into the corresponding embedding matrix, query matrix, and reference matrix: ， ； ， ； ， ； in, and These represent the conversion of the left hemisphere enhancement features and the right hemisphere enhancement features into their corresponding embedding matrices, respectively. and These represent the weight matrices corresponding to the embedding matrices of the left and right hemispheres, respectively. and These represent the conversion of the left hemisphere enhanced features and the right hemisphere enhanced features into corresponding query matrices, respectively. and These represent the weight matrices corresponding to the query matrices in the left and right hemispheres, respectively. and These represent the transformations of the left hemisphere enhancement features and the right hemisphere enhancement features into their corresponding reference matrices, respectively. and These represent the weight matrices corresponding to the reference matrices for the left and right hemispheres, respectively. Calculate the cross-attention mechanism based on all the reference matrices and all the query matrices: ； ； in, and These represent the cross-attention mechanisms targeting the left and right hemispheres, respectively. Represents the normalized exponential function, Indicates the dimensions of the query matrix and parameter matrix. For transpose, Indicates to Transpose. Indicates to Transpose; The embedding matrix and cross-attention mechanism corresponding to the left hemisphere are input into the first gating filter mechanism, and the embedding matrix and cross-attention mechanism corresponding to the right hemisphere are input into the second gating filter mechanism, which outputs the training features of the left hemisphere and the training features of the right hemisphere, respectively. ； ； in, and These represent the results after the first linear layer. and , Represents the hyperbolic tangent function. This represents the second activation function. and These represent the results after the second linear layer. and , and These represent the gating scores calculated for the left and right hemispheres, respectively. This represents the third activation function. and These represent the training characteristics of the left hemisphere and the right hemisphere, respectively. , , , , , , , , and These represent different linear layers.

7. The fine-grained brain age prediction method according to claim 6, characterized in that, The process of using the cross-hemispheric attention model to iteratively process all the training features in multiple rounds, and inputting all the feature representations into a multilayer perceptron for prediction, outputting the predicted brain age of the target object in different regions, specifically includes: The training features of the left hemisphere and the training features of the right hemisphere are input into the next spatial mixer and channel mixer in the cross-hemispherical attention model for iterative processing until all processes are completed, and the final prediction data of each triangular grid is obtained. All the final prediction data are input into a multilayer perceptron for prediction, and the predicted brain age of each triangular grid of the target object is output: ； in, Indicates the first Predicted brain age using a triangular grid. Indicates a linear layer. Indicates the first The final predicted data.

8. A fine-grained brain age prediction system, characterized in that, The fine-grained brain age prediction system includes: The image preprocessing module is used to acquire the initial MRI image of the target object's brain, segment the initial MRI image to obtain the tissue types of different regions, and calculate the curvature feature map, cortical thickness feature map and sulcus depth feature map based on all the tissue types. The spatial mixing module is used to resample the brain onto a polyhedron based on the curvature feature map, the cortical thickness feature map, and the sulcus depth feature map to obtain multiple feature representations, and to convert all the feature representations into corresponding embedded features; The channel mixing module is used to input all the embedded features into the constructed cross-hemispherical attention model. The cross-hemispherical attention model converts all the embedded features into multiple global representations and concatenates all the global representations with the multiple embedded features to obtain multiple interactive features. The dynamic adaptive bias module is used to uniformly divide all the interaction features and then perform enhancement processing on each of them to obtain the enhanced features corresponding to each interaction feature. All the enhanced features are converted into multiple matrices, and multiple training features are generated based on all the matrices through a cross-attention mechanism and a gating filtering mechanism. The brain age prediction module is used to perform multiple rounds of iterative processing on all the training features using the cross-hemispheric attention model, and input all the feature representations into a multilayer perceptron for prediction, and output the predicted brain age of the target object in different regions.

9. A terminal, characterized in that, The terminal includes: a memory, a processor, and a fine-grained brain age prediction program stored in the memory and executable on the processor, wherein the fine-grained brain age prediction program, when executed by the processor, implements the steps of the fine-grained brain age prediction method as described in any one of claims 1-7.

10. A storage medium, characterized in that, The storage medium stores a fine-grained brain age prediction program, which, when executed by a processor, implements the steps of the fine-grained brain age prediction method as described in any one of claims 1-7.

Images