Imaging-based brain network to morphological brain network prediction methods, systems, and devices

Abstract

The present application relates to the technical field of medical image generation, in particular to a method, system and device for predicting a morphological brain network from an image-based brain network, to solve the technical problem of poor prediction effect of the morphological brain network in the prior art, the present application first extracts the image space features of the image-based brain network and the brain region image-based features of T1-weighted imaging, extracts the multi-level structure information from the local to the global of the brain region, and then combines the residual connection graph convolution and the multi-layer perception mechanism to optimize the feature representation, and obtains the image space features; at the same time, the Euclidean space feature extraction is used to supplement the spatial position information of the brain region, and the image space features are fused in space, so as to fully capture the topological structure, symmetry characteristics and spatial distribution law of the morphological brain network, and effectively improve the accuracy, symmetry and robustness of the image-based brain network in predicting the morphological brain network based on T1-weighted imaging.

Description

Technical Field

[0001] This invention relates to the field of medical image generation technology, specifically to methods, systems, and devices for predicting brain networks from radiomics to morphological brain networks. Background Technology

[0002] In neuroscience research, morphological brain networks have been widely used to explore changes in structural connectivity during brain development, aging, and mental and neurological diseases, providing an important network perspective on brain structural integration and differentiation.

[0003] In constructing morphological brain networks, morphological feature extraction is typically very time-consuming. Extracting morphological features such as cortical thickness, gray matter volume, and surface area from a single subject usually takes 6 to 10 hours. Such high computational costs not only extend the data processing cycle but also greatly limit the promotion and popularization of morphological brain networks in large-scale samples and clinical applications. Brain network modeling methods can significantly reduce the time required for morphological brain network feature representation. However, existing brain network modeling methods often ignore the fact that the brain is a typical Euclidean-non-Euclidean mixed spatial structure. Relying solely on spatial structural information fails to fully capture the potential structural patterns of the brain network, resulting in poor predictive performance and weak expressive power of morphological brain networks. Summary of the Invention

[0004] The purpose of this invention is to provide a method, system, and device for predicting brain networks from radiomics to morphological brain networks.

[0005] The technical solution of this invention is as follows:

[0006] A method for predicting brain networks from radiomics to morphological brain networks, characterized by comprising the following operations:

[0007] S1. Obtain brain region radiomics features, radiomics brain networks, and morphological brain networks from multiple T1-weighted imaging to form a brain network dataset. Use the brain network dataset to train a prediction network until the sum of the network loss between the predicted morphological brain networks and the symmetric structure loss of the predicted morphological brain networks is less than the loss threshold, then stop training and obtain the trained prediction network.

[0008] In the process of predicting the network, brain region radiomics features and radiomics brain networks are processed by graph spatial feature extraction to obtain graph spatial features; the radiomics brain network is processed by Euclidean spatial feature extraction to obtain Euclidean spatial features; the graph spatial features and Euclidean spatial features are processed by spatial fusion to obtain the predictive morphological brain network.

[0009] The graph space feature extraction process is as follows: brain region radiomics features and radiomics brain networks are processed by graph convolution at different scales to obtain multi-depth radiomics brain network features; multi-depth radiomics brain network features are then processed by fusion based on symmetry-guided attention to obtain radiomics brain network fusion features; radiomics brain network fusion features are then processed by graph convolution, graph convolution based on residual connections, and multilayer perception mechanism to obtain graph space features.

[0010] S2. The radiomics brain network of T1-weighted imaging to be processed is processed by the training prediction network to obtain the morphological brain network prediction results.

[0011] Brain region radiomics features include intensity and texture features of the brain region; intensity features include mean, standard deviation, minimum and maximum gray values ​​of the brain region.

[0012] The Euclidean space feature extraction operation is implemented using the following formula:

[0013] ,

[0014] ,

[0015] ,

[0016] These are the upper triangular elements of the radiomics brain network. H It features European-style spatial characteristics. , These are the first nonlinear characteristic and the second nonlinear characteristic, respectively. , , The first nonlinear weight, the second nonlinear weight, and the third nonlinear weight are, , , These are the first nonlinear bias, the second nonlinear bias, and the third nonlinear bias, respectively. This is the Sigmoid function.

[0017] The graph convolution operation at different scales is implemented using the following formula:

[0018] ,

[0019] ,

[0020] ,

[0021] ,

[0022] ,

[0023] These are radiomic features of brain regions. Linear features of brain region radiomics. For linear processing, For radiomics brain networks, the adjacency matrix is... This represents the edge representation feature matrix of a radiomics brain network. The angle matrix of radiomics brain networks. , , The brain network features are represented by the first deep imaging group, the second deep imaging group, and the third deep imaging group. , , These are the first-scale weights, the second-scale weights, and the third-scale weights, respectively. It is the ReLU activation function. This is the Dropout function.

[0024] The fusion processing based on symmetry-guided attention involves: obtaining the basic attention weights and brain region symmetry weights of multi-depth radiomics brain network features, respectively; after aggregation, performing a weighted summation with their respective deep radiomics brain network features to obtain the radiomics brain network fusion features; the basic attention weights of the deep radiomics brain network features are obtained by linear mapping and Sigmoid function processing of the deep radiomics brain network features; the brain region symmetry weights of the deep radiomics brain network features are obtained by summing the basic attention weights of the brain regions with the basic attention weights of the corresponding symmetrical brain regions.

[0025] The predicted network loss between morphological brain networks is calculated using the following formula:

[0026] ,

[0027] To predict network loss between morphological brain networks, To predict morphological brain networks, For morphological brain networks, To predict morphological brain networks and their covariance, , These are the predicted variances of morphological brain networks and the variances of morphological brain networks, respectively. To predict the paradigm processing results of morphological brain networks.

[0028] The loss of symmetric structure in morphological brain networks is calculated using the following formula:

[0029] ,

[0030] To predict the loss of symmetry structure in morphological brain networks, P represents the total number of symmetric brain regions. It is a collection of symmetrical brain regions. To predict brain regions in morphological brain networks i and brain regions j The strength of brain region connectivity, To predict brain regions in morphological brain networks j and brain regions i The strength of brain region connectivity, This is handled in the second normal form.

[0031] The spatial fusion process involves extracting the upper triangular elements of the graph space features, expanding them into a vector, and then fusioning them with the Euclidean space features using weighted methods to restore them to a symmetric matrix, thereby obtaining the predicted morphological brain network.

[0032] A predictive system for radiomics-to-morphological brain networks, used to implement the aforementioned predictive method for radiomics-to-morphological brain networks, includes:

[0033] A training prediction network generation module is used to acquire multiple T1-weighted imaging brain region radiomics features, radiomics brain networks, and morphological brain networks, forming a brain network dataset. Using this dataset, the prediction network is trained until the sum of the network loss between predicted morphological brain networks and the loss of the symmetric structure of predicted morphological brain networks is less than a loss threshold, at which point training stops, resulting in a trained prediction network. During the prediction network processing, the brain region radiomics features and the radiomics brain networks undergo graph space feature extraction to obtain graph space features; the radiomics brain networks are then processed using Euclidean space... The graph spatial features are extracted to obtain Euclidean spatial features. The graph spatial features and Euclidean spatial features are then spatially fused to obtain the predicted morphological brain network. The graph spatial feature extraction process is as follows: brain region radiomics features and radiomics brain networks are processed by graph convolution at different scales to obtain multi-depth radiomics brain network features. The multi-depth radiomics brain network features are then fused by symmetry-guided attention to obtain radiomics brain network fusion features. The radiomics brain network fusion features are then processed by graph convolution, graph convolution based on residual connections, and multilayer perception mechanism to obtain graph spatial features.

[0034] The morphological brain network prediction result generation module is used to process the radiomics brain network of T1-weighted imaging to be processed, and obtain the morphological brain network prediction result after training the prediction network.

[0035] A device for predicting radiomics-to-morphological brain networks includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the aforementioned method for predicting radiomics-to-morphological brain networks.

[0036] A computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the above-described method for predicting from radiomics brain networks to morphological brain networks.

[0037] The beneficial effects of this invention are as follows:

[0038] This invention provides a method for predicting morphological brain networks from radiomics brain networks. First, graph spatial features are extracted from T1-weighted imaging brain region radiomics features and the radiomics brain network itself. This extracts multi-level structural information from local to global brain regions. Features related to the core symmetry attributes of the morphological brain network are then enhanced through symmetry-guided attention fusion. Further feature representation is optimized using graph convolution with residual connections and multilayer perceptron mechanisms to obtain graph spatial features. Simultaneously, Euclidean spatial features are used to supplement spatial location information of brain regions, resulting in Euclidean spatial features. Finally, graph spatial features and Euclidean spatial features are spatially fused, integrating key clues of structural and spatial dimensions to comprehensively capture the topological structure, symmetry characteristics, and spatial distribution patterns of the morphological brain network. This effectively improves the accuracy, symmetry, and robustness of predicting morphological brain networks using T1-weighted imaging radiomics brain networks. Attached Figure Description

[0039] The solutions and advantages of this application will become clear to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of the invention.

[0040] In the attached diagram:

[0041] Figure 1 This is a flowchart illustrating the prediction method in this embodiment.

[0042] Figure 2 This is a comparison diagram of the predicted morphological brain network and the real morphological brain network in the embodiment. Figure 2 In the diagram, (a) represents the actual morphological brain network, and (b) represents the predicted morphological brain network. Detailed Implementation

[0043] Exemplary embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings.

[0044] A method for predicting brain networks from radiomics to morphological brain networks, see [link to relevant documentation]. Figure 1 This includes the following operations:

[0045] S1. Obtain brain region radiomics features, radiomics brain networks, and morphological brain networks from multiple T1-weighted imaging to form a brain network dataset. Use the brain network dataset to train a prediction network until the sum of the network loss between the predicted morphological brain networks and the symmetric structure loss of the predicted morphological brain networks is less than the loss threshold, then stop training and obtain the trained prediction network.

[0046] S2. The radiomics brain network of T1-weighted imaging to be processed is processed by the training prediction network to obtain the morphological brain network prediction results.

[0047] The specific steps are detailed below.

[0048] S1. Obtain brain region radiomics features, radiomics brain networks, and morphological brain networks from multiple T1-weighted imaging studies to form a brain network dataset. Using the brain network dataset, train the prediction network until the sum of the network loss between the predicted morphological brain networks and the symmetric structure loss of the predicted morphological brain networks is less than the loss threshold, then stop training and obtain the trained prediction network.

[0049] First, we acquire multiple T1-weighted imaging brain region radiomics features, radiomics brain networks, and morphological brain networks. A T1-weighted imaging brain region radiomics feature and radiomics brain network, along with the corresponding morphological brain network, form a training sample pair. All training sample pairs form a brain network dataset.

[0050] The brain region radiomics features obtained from T1-weighted imaging are based on brain region texture and intensity features. Specifically, T1-weighted imaging is subjected to N4 bias field correction, linear registration, and nonlinear registration to the MNI152 standard space. Twenty-five radiomics features are extracted for each brain region, including intensity and texture features. The intensity features include the (registered) mean, standard deviation, minimum, and maximum gray values ​​of the brain region.

[0051] The method for obtaining the radiomics brain network of T1-weighted imaging is as follows: T1-weighted imaging is subjected to N4 bias field correction, linear registration and nonlinear registration to the MNI152 standard space, the whole brain is divided into regions based on the Desikan-Killiany brain map, each brain region is divided into a node, and the features of each node are its (brain region) radiomics features. Pearson correlation coefficients are calculated based on the (brain region) radiomics features between brain regions (between nodes) and used as edges between nodes to obtain the radiomics brain network reflecting the connectivity between brain regions.

[0052] The method for obtaining the (real) morphological brain network from T1-weighted imaging is as follows: T1-weighted imaging is subjected to N4 bias field correction, linear registration and nonlinear registration to the MNI152 standard space, the whole brain is divided into regions based on the Desikan-Killiany brain map, and the morphological feature vector of each brain region is obtained. The similarity between the morphological feature vectors of any two brain regions is calculated using the Pearson correlation coefficient to form a symmetric similarity matrix. Each element in this matrix represents the degree of morphological similarity between the corresponding brain regions, thus forming a weighted undirected network as the morphological brain network.

[0053] Then, using the brain network dataset, the prediction network is trained until the sum of the network loss between the predicted morphological brain networks and the symmetric structure loss of the predicted morphological brain networks is less than the loss threshold. The model converges, training stops, and the trained prediction network is obtained.

[0054] In the process of predicting the network, brain region radiomics features and radiomics brain networks are processed by graph spatial feature extraction to obtain graph spatial features; radiomics brain networks are processed by Euclidean spatial feature extraction to obtain Euclidean spatial features; graph spatial features and Euclidean spatial features are processed by spatial fusion to obtain the predictive morphological brain network.

[0055] The specific steps for the above graph space feature extraction process are as follows.

[0056] Step 1: Brain region radiomics features. The radiomics brain network is processed by graph convolution at different scales to extract features of local, global, multi-level, complementary, and adaptive morphological characteristics of brain regions, resulting in multi-depth radiomics brain network features, which provide a comprehensive and accurate information foundation for subsequent prediction.

[0057] The graph convolution operation at different scales is implemented using the following formula:

[0058] ,

[0059] ,

[0060] ,

[0061] ,

[0062] ,

[0063] These are radiomic features of brain regions. Linear features of brain region radiomics. For linear processing, For radiomics brain networks, the adjacency matrix is... The edge representation feature matrix of a radiomics brain network is obtained by transforming the adjacency matrix of the radiomics brain network. Obtained by graph convolution operation defined by Laplace transform. The angle matrix of radiomics brain networks. This section focuses on the first-depth imaging-based brain network features, highlighting local connectivity characteristics of adjacent brain regions. To identify the second-depth radiomics brain network features, we focused on the transhemispheric connections between corresponding brain regions in the left and right hemispheres. To explore the characteristics of the third-depth radiomics brain network, we focus on the "small-world" topological properties of the whole-brain network. , , These are the first-scale weights, the second-scale weights, and the third-scale weights, respectively. It is the ReLU activation function. This is the Dropout function.

[0064] Step 2: Multi-depth radiomics brain network features are fused using symmetry-guided attention to enhance the symmetry core, coordinate feature conflicts, filter redundancy, and dynamically adapt to individual differences, making the fused features closer to the essential attributes of morphological brain networks, thus obtaining radiomics brain network fusion features.

[0065] The specific operation of the fusion processing based on symmetry-guided attention is as follows: the basic attention weights and brain region symmetry weights of the multi-depth radiomics brain network features are obtained respectively, and after aggregation processing, they are weighted and summed with the respective deep radiomics brain network features to obtain the radiomics brain network fusion features.

[0066] The basic attention weights of the aforementioned deep radiomics brain network features are obtained by linearly mapping the deep radiomics brain network features and processing them with the Sigmoid function, specifically through the following formula:

[0067] ,

[0068] Based on attention weights, For the Sigmoid function, The features of deep imaging brain networks (first deep imaging brain network features, or second deep imaging brain network features, or third deep imaging brain network features). is the linear mapping constant.

[0069] The brain region symmetry weights of deep radiomics brain network features are obtained by summing the basic attention weights of the brain regions with the basic attention weights of the corresponding symmetrical brain regions. i The symmetry weights of individual brain regions can be calculated using the following formula:

[0070] ,

[0071] For the first i Symmetrical weights of individual brain regions For the first i Symmetrical brain regions corresponding to each brain region j Basic attention weights It is a collection of symmetrical brain regions, consisting of brain regions i and brain regions j composition.

[0072] The above aggregation process can be implemented using the following formula:

[0073]

[0074] To update the weights, For the clip function, The polymerization coefficient is denoted as .

[0075] Step 3: The radiomics brain network fusion features are processed by graph convolution, graph convolution based on residual connections (preferably twice), and multilayer sensing mechanism to obtain the radiomics brain network fusion features. Among them, the multilayer sensing mechanism consists of one input layer, two hidden layers, and one output layer. Each hidden layer is followed by a Dropout function and a ReLU function to prevent overfitting and make the prediction network more robust.

[0076] The above Euclidean space feature extraction process involves first extracting the upper triangular elements of the radiomics brain network and expanding them into a vector. This vector is then subjected to a nonlinear transformation through three fully connected layers to model the nonlinear interactions between brain regions and capture potential higher-order structural information in Euclidean space, thereby improving the prediction accuracy of the morphological brain network. This Euclidean space feature extraction process can be implemented using the following formula:

[0077] ,

[0078] ,

[0079] ,

[0080] It features European-style spatial characteristics. These are the upper triangular elements of the radiomics brain network. , These are the first nonlinear characteristic and the second nonlinear characteristic, respectively. , , The first nonlinear weight, the second nonlinear weight, and the third nonlinear weight are, , , These are respectively the first nonlinear paranoia, the second nonlinear paranoia, and the third nonlinear paranoia. This is the Sigmoid function.

[0081] The spatial fusion process described above is as follows: Extract the upper triangular elements of the graph space features, expand them into a vector, and then fuse them with the Euclidean space features using a weighted method to restore them to a symmetric matrix, thus obtaining the predicted morphological brain network. The calculation formula is as follows:

[0082] ,

[0083] To predict morphological brain networks, , These are graph space features (upper triangular elements) and Euclidean space features, respectively. , These are graph space weights and Euclidean space weights, respectively. This is a function for restoring symmetric matrices.

[0084] The network loss between the predicted morphological brain networks is calculated using the following formula:

[0085] ,

[0086] To predict network loss between morphological brain networks, To predict morphological brain networks, For morphological brain networks, To predict morphological brain networks and their covariance, , These are the predicted variances of morphological brain networks and the variances of morphological brain networks, respectively. To predict the paradigm processing results of morphological brain networks.

[0087] The loss of symmetric structure in morphological brain networks is calculated using the following formula:

[0088] ,

[0089] To predict the loss of symmetric structures in morphological brain networks, P represents the total number of symmetric brain regions. A collection of symmetrical brain regions (composed of brain regions) i and brain regions j composition), To predict brain regions in morphological brain networks i and brain regions j The strength of brain region connectivity, To predict brain regions in morphological brain networks j and brain regions i The strength of brain region connectivity, This is handled in the second normal form.

[0090] S2. The radiomics brain network of T1-weighted imaging to be processed is processed by the training prediction network to obtain the morphological brain network prediction results.

[0091] The brain region radiomics features and radiomics brain networks from T1-weighted imaging to be processed are then processed by a training prediction network to obtain morphological brain network prediction results.

[0092] To verify the effectiveness of the morphological brain network generation method in this embodiment, an experiment was conducted. In the experiment, the dataset was randomly divided into training, validation, and test sets in a 6:2:2 ratio. The experimental development environment used was PyTorch 2.6.0 on an NVIDIA RTX 2080 graphics processor. The training process consisted of 100 epochs, using the Adam optimizer, with an initial learning rate of... The batch size was set to 16. The experimental environment and specific settings are shown in Table 1. Experimental results can be found in [reference needed]. Figure 2 According to Table 2, the experimental results show that the mean squared error between the predicted morphological brain network connection distribution and the connection values ​​of the actual morphological brain network generated in this embodiment reaches 0.062, and the Pearson correlation coefficient is 0.784. Furthermore, from the visualization results of individual results (see Table 2), the mean squared error between the predicted morphological brain network connection distribution and the actual morphological brain network connection values ​​reaches 0.062, and the mean squared error between the predicted morphological brain network connection distribution and the actual morphological brain network connection values ​​is 0.784. Figure 2 It can also be seen that the predicted morphological brain network generated in this embodiment is very close to the real morphological brain network, and the strength of the topology and the distribution of connection values ​​can be clearly seen.

[0093] Table 1 Summary of Experimental Parameters

[0094]

[0095] Table 2 Summary of Experimental Results Parameters

[0096]

[0097] This embodiment also provides a predictive system for radiomics-to-morphological brain networks, used to implement the above-mentioned predictive method for radiomics-to-morphological brain networks, including:

[0098] A training prediction network generation module is used to acquire multiple T1-weighted imaging brain region radiomics features, radiomics brain networks, and morphological brain networks, forming a brain network dataset. Using this dataset, the prediction network is trained until the sum of the network loss between predicted morphological brain networks and the loss of the symmetric structure of predicted morphological brain networks is less than a loss threshold, at which point training stops, resulting in a trained prediction network. During the prediction network processing, the brain region radiomics features and the radiomics brain networks undergo graph space feature extraction to obtain graph space features; the radiomics brain networks are then processed using Euclidean space... The graph spatial features are extracted to obtain Euclidean spatial features. The graph spatial features and Euclidean spatial features are then spatially fused to obtain the predicted morphological brain network. The graph spatial feature extraction process is as follows: brain region radiomics features and radiomics brain networks are processed by graph convolution at different scales to obtain multi-depth radiomics brain network features. The multi-depth radiomics brain network features are then fused by symmetry-guided attention to obtain radiomics brain network fusion features. The radiomics brain network fusion features are then processed by graph convolution, graph convolution based on residual connections, and multilayer perception mechanism to obtain graph spatial features.

[0099] The morphological brain network prediction result generation module is used to process the radiomics brain network of T1-weighted imaging to be processed, and obtain the morphological brain network prediction result after training the prediction network.

[0100] This embodiment also provides a device for predicting radiomics brain networks to morphological brain networks, including a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the above-described method for predicting radiomics brain networks to morphological brain networks.

[0101] This embodiment also provides a computer-readable storage medium for storing a computer program, wherein the computer program, when executed by a processor, implements the above-described method for predicting from radiomics brain networks to morphological brain networks.

[0102] This embodiment provides a method for predicting morphological brain networks from radiomics brain networks. First, graph spatial features are extracted from T1-weighted imaging brain region radiomics features and radiomics brain networks to extract multi-level structural information from local to global brain regions. Features related to the core symmetry attributes of morphological brain networks are enhanced through symmetry-guided attention fusion. Then, graph convolution with residual connections and multilayer perception mechanisms are combined to optimize feature representation and obtain graph spatial features. Simultaneously, Euclidean spatial features are used to supplement the spatial location information of brain regions to obtain Euclidean spatial features. Finally, graph spatial features and Euclidean spatial features are spatially fused to integrate key clues of structural and spatial dimensions, comprehensively capturing the topological structure, symmetry characteristics, and spatial distribution patterns of morphological brain networks. This effectively improves the accuracy, symmetry, and robustness of predicting morphological brain networks based on T1-weighted imaging radiomics brain networks.

[0103] This embodiment provides a method for predicting brain networks from radiomics to morphological brain networks. It enables the artificial intelligence synthesis of brain morphological networks without extracting morphological features such as brain gray matter volume, which greatly reduces the generation time and construction cost of morphological brain networks, while also enabling the synthesis of high-quality morphological brain networks.

Claims

1. A method for predicting the transformation from radiomics brain networks to morphological brain networks, characterized in that, This includes the following operations: S1. Obtain brain region radiomics features, radiomics brain networks, and morphological brain networks from multiple T1-weighted imaging to form a brain network dataset. Use the brain network dataset to train a prediction network until the sum of the network loss between the predicted morphological brain networks and the symmetric structure loss of the predicted morphological brain networks is less than the loss threshold, then stop training and obtain the trained prediction network. In the process of predicting the network, brain region radiomics features and radiomics brain networks are processed by graph spatial feature extraction to obtain graph spatial features; the radiomics brain network is processed by Euclidean spatial feature extraction to obtain Euclidean spatial features; the graph spatial features and Euclidean spatial features are processed by spatial fusion to obtain the predictive morphological brain network. The graph space feature extraction process is as follows: brain region radiomics features and radiomics brain networks are processed by graph convolution at different scales to obtain multi-depth radiomics brain network features; multi-depth radiomics brain network features are then processed by fusion based on symmetry-guided attention to obtain radiomics brain network fusion features; radiomics brain network fusion features are then processed by graph convolution, graph convolution based on residual connections, and multilayer perception mechanism to obtain graph space features. S2. The radiomics brain network of T1-weighted imaging to be processed is processed by the training prediction network to obtain the morphological brain network prediction results.

2. The method for predicting from radiomics brain networks to morphological brain networks according to claim 1, characterized in that, Brain region radiomics features include intensity and texture features of the brain region; intensity features include mean, standard deviation, minimum and maximum gray values ​​of the brain region.

3. The method for predicting the transformation from radiomics brain networks to morphological brain networks according to claim 1, characterized in that, The Euclidean space feature extraction operation is implemented using the following formula: ， ， ， H It features European-style spatial characteristics. These are the upper triangular elements of the radiomics brain network. , These are the first nonlinear characteristic and the second nonlinear characteristic, respectively. , , The first nonlinear weight, the second nonlinear weight, and the third nonlinear weight are, , , These are the first nonlinear bias, the second nonlinear bias, and the third nonlinear bias, respectively. This is the Sigmoid function.

4. The method for predicting from radiomics brain networks to morphological brain networks according to claim 1, characterized in that, The graph convolution operation at different scales is implemented using the following formula: ， ， ， ， ， These are radiomic features of brain regions. Linear features of brain region radiomics. For linear processing, For radiomics brain networks, the adjacency matrix is... This represents the edge representation feature matrix of a radiomics brain network. The angle matrix of radiomics brain networks. , , The brain network features are represented by the first deep imaging group, the second deep imaging group, and the third deep imaging group. , , These are the first-scale weights, the second-scale weights, and the third-scale weights, respectively. It is the ReLU activation function. This is the Dropout function.

5. The method for predicting from radiomics brain networks to morphological brain networks according to claim 1, characterized in that, The fusion processing based on symmetry-guided attention is as follows: The basic attention weights and brain region symmetry weights of the multi-depth radiomics brain network features were obtained separately. After aggregation, they were weighted and summed with the respective deep radiomics brain network features to obtain the radiomics brain network fusion features. The basic attention weights of deep radiomics brain network features are obtained by linearly mapping and processing the deep radiomics brain network features using the Sigmoid function. The brain region symmetry weights of deep radiomics brain network features are obtained by summing the basic attention weights of the brain regions with the basic attention weights of the corresponding symmetrical brain regions.

6. The method for predicting from radiomics brain networks to morphological brain networks according to claim 1, characterized in that, The predicted network loss between morphological brain networks is calculated using the following formula: ， To predict network loss between morphological brain networks, To predict morphological brain networks, For morphological brain networks, To predict morphological brain networks and their covariance, , These are the predicted variances of morphological brain networks and the variances of morphological brain networks, respectively. To predict the paradigm processing results of morphological brain networks; The loss of symmetric structure in morphological brain networks is calculated using the following formula: ， To predict the loss of symmetry structure in morphological brain networks, P represents the total number of symmetric brain regions. It is a collection of symmetrical brain regions. To predict brain regions in morphological brain networks i and brain regions j The strength of brain region connectivity, To predict brain regions in morphological brain networks j and brain regions i The strength of brain region connectivity, This is handled in the second normal form.

7. The method for predicting from radiomics brain networks to morphological brain networks according to claim 1, characterized in that, The spatial fusion process involves extracting the upper triangular elements of the graph space features, expanding them into a vector, and then fusioning them with the Euclidean space features using weighted methods to restore them to a symmetric matrix, thereby obtaining the predicted morphological brain network.

8. A predictive system for converting radiomics brain networks into morphological brain networks, characterized in that, The method for predicting the transformation from radiomics brain networks to morphological brain networks as described in claim 1, characterized in that it comprises: A training prediction network generation module is used to acquire multiple T1-weighted imaging brain region radiomics features, radiomics brain networks, and morphological brain networks, forming a brain network dataset. Using this dataset, the prediction network is trained until the sum of the network loss between predicted morphological brain networks and the loss of the symmetric structure of predicted morphological brain networks is less than a loss threshold, at which point training stops, resulting in a trained prediction network. During the prediction network processing, the brain region radiomics features and the radiomics brain networks undergo graph space feature extraction to obtain graph space features; the radiomics brain networks are then processed using Euclidean space... The graph spatial features are extracted to obtain Euclidean spatial features. The graph spatial features and Euclidean spatial features are then fused to obtain the predicted morphological brain network. The graph spatial feature extraction process is as follows: brain region radiomics features and radiomics brain networks are processed by graph convolution at different scales to obtain multi-depth radiomics brain network features. The multi-depth radiomics brain network features are then fused by symmetry-guided attention to obtain radiomics brain network fusion features. The radiomics brain network fusion features are then processed by graph convolution, graph convolution based on residual connections, and multilayer perception mechanism to obtain graph spatial features. The morphological brain network prediction result generation module is used to process the radiomics brain network of T1-weighted imaging to be processed, and obtain the morphological brain network prediction result after training the prediction network.

9. A predictive device for converting radiomics brain networks into morphological brain networks, characterized in that, It includes a processor and a memory, wherein the processor executes a computer program stored in the memory to implement the method for predicting radiomics brain networks to morphological brain networks as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, Used to store a computer program, wherein the computer program, when executed by a processor, implements the method for predicting from radiomics brain networks to morphological brain networks as described in any one of claims 1-7.

Images