Brain disease discrimination deep learning system based on meta analysis and individualized attention mixed guidance

Abstract

The invention discloses a brain disease discrimination deep learning system based on meta analysis and individualized attention mixed guidance. The system comprises a ReHo feature extraction module; the meta analysis atlas construction module and the meta atlas integration analysis module are used for establishing a threshold screening mechanism and a spatial topological feature enhancement strategy for the neural image feature atlas to obtain a standardized meta analysis atlas; the brain disease discrimination construction training module comprises an STA network and an MAH network, the STA network is used for generating an attention map for each input brain image and directly using a classification label as weak supervision positioning guidance, and the MAH network is used for fusing a meta analysis map, a ReHo index and the attention map and outputting a brain disease discrimination result. According to the method, an own attention map is generated for each subject through the STA network, and the constructed global branch is fused into the deep learning model to provide individual level guidance, so that high-accuracy discrimination of brain diseases is completed.

Description

Technical Field

[0001] The present invention belongs to the field of image processing, and specifically relates to a deep learning system for brain disease discrimination based on a hybrid guidance of meta-analysis and individual attention. Background Art

[0002] Functional Magnetic Resonance Imaging (fMRI) technology is a common imaging tool for researchers to study human brain diseases. It is a non-invasive imaging technology used to study the functional interaction patterns of the human brain and can effectively detect brain tissue and its related functional activation patterns. fMRI has the advantages of high spatial resolution and non-invasiveness, and can well evaluate brain diseases without obvious structural lesions.

[0003] With the great success of deep learning technology in computer vision tasks, Convolutional Neural Networks (CNNs) and Transformers have shown promising applications in the field of neuroimaging research due to their powerful ability to learn representative and predictive features from large-scale data in a task-oriented manner. Deep learning models based on CNNs have been applied to the automatic diagnosis of mental diseases and can play an important role in diagnosis by reducing subjectivity and improving the repeatability of diagnosis. Despite decades of psychiatric research, the discrimination of mental diseases remains limited. The small scale of medical datasets is still a challenge in obtaining a satisfactory deep learning model, especially when considering the discrimination problem of brain diseases, and disease heterogeneity exacerbates this problem.

[0004] As a parallel research branch of data-driven models, a large number of in-depth statistical studies have been carried out on mental diseases based on small-scale data, and valuable neural circuit priors of abnormal brain functions related to mental diseases have been obtained. However, the heterogeneity of mental disease patients, that is, the individual differences are significant, and the group-based neural circuit priors are difficult to accurately fit a single subject. Therefore, it is an urgent problem to integrate an individual attention mechanism on the basis of neural circuit prior information to provide hybrid guidance at the individual level and the group level for the model, and to improve the robustness and generalization of the model through the collaborative optimization mechanism of the meta-analysis atlas and the attention atlas. Summary of the Invention

[0005] Aiming at the problems existing in the prior art, the present invention provides a deep learning system for brain disease discrimination based on a hybrid guidance of meta-analysis and individual attention, which can improve the robustness and generalization of the model through the collaborative optimization mechanism of the meta-analysis atlas and the attention atlas.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: a deep learning system for brain disease discrimination based on a hybrid guidance of meta-analysis and individual attention,

[0007] including: a ReHo feature extraction module that obtains the ReHo index based on the local consistency of magnetic resonance imaging from brain images;

[0008] a meta-analysis atlas construction module for comparing the significantly different voxel functions and structures and spatially convergent brain regions between patients with depression and healthy subjects to obtain a neuroimaging feature atlas generated by meta-analysis;

[0009] a meta-analysis atlas integration and analysis module for establishing a threshold screening mechanism and a spatial topological feature enhancement strategy for the neuroimaging feature atlas generated by meta-analysis to obtain a standardized meta-analysis atlas M;

[0010] a brain disease discrimination construction and training module including a STA network and a MAH network. The STA network is used to generate an attention map for each input brain image and directly uses the classification label as weak supervision for localization guidance. The MAH network is used to fuse the meta-analysis atlas M, the ReHo index, and the attention map and output the brain disease discrimination result.

[0011] Further, the aforementioned ReHo index is calculated as follows:

[0012]

[0013] where the functional magnetic resonance imaging (fMRI) time series contains N time nodes, N being the total number of scanning phases. For each specific time point i, the signal intensities of all voxels in the whole brain at this moment are sorted in ascending order through statistical methods to generate a sorting variable R i , represents the arithmetic mean of the R i sorting values corresponding to all time points, and K is a fixed value corresponding to the number of adjacent voxels in the three-dimensional spatial neighborhood analysis.

[0014] Further, the aforementioned sorting variable R i is:

[0015]

[0016] where the sorting of the signal value collected by voxel j at time node i in the time series is r i,j .

[0017] Further, the arithmetic mean of the aforementioned sorting variable R i sorting values is:

[0018]

[0019] Furthermore, the aforementioned meta-graph integration analysis module is configured to perform the following actions:

[0020] S101. Perform dynamic weight d weight quantitative calculation on each voxel unit as follows:

[0021] d weight = log(PR(x));

[0022] where each spatial unit x corresponds to the bibliometric parameter PR(x), and this parameter quantitatively characterizes the attention index of this brain region in historical research, with a value range of 0 - 1. PR(x) = 0 indicates that this coordinate unit has not been reported in existing literature;

[0023] S102. Define the mathematical representation of the meta-graph as:

[0024]

[0025] where d is the Euclidean distance from the center of the spatial unit to the activation peak point, and σ is the spatial decay coefficient of the Gaussian kernel function; S103. Construct the spatial unit optimization index M n as follows:

[0026]

[0027] S104. Introduce an empirical threshold to implement feature screening. The units smaller than the empirical threshold are regarded as low-confidence signals and removed, and finally the optimized weight parameter M is generated:

[0028]

[0029] Construct the standardized meta-analysis graph M in this way as the input of the deep learning model.

[0030] Furthermore, the aforementioned STA network includes a sequentially connected convolutional layer, a linear projection layer, a first STF module, a second STF module, a third STF module, an average pooling layer, and a classification layer; the STF module includes a LayerNorm layer, an MSA layer, and an MLP layer. The convolutional layer serves as the input end of the STA network, receives the ReHo index, and calculates the attention map as follows:

[0031]

[0032] Classification score s c :

[0033]

[0034] Aggregation of the quantified attention maps of different categories:

[0035] A(x,y,z) = ∑ c s c A c (x,y,z),

[0036] where F m (x,y,z) is the feature map output by the STA, is the parameter matrix of the FC layer after STA, A c is the attention map of one channel, A is the final attention map, s c is the weight of each channel attention map.

[0037] Furthermore, the aforementioned deep learning system for brain disease discrimination guided by the hybrid of meta-analysis and individual attention takes the ReHo index, the meta-analysis map, and the attention map generated by the STA network as inputs, the brain disease discrimination result as the output, and constructs the MAH network;

[0038] The ReHo index, the meta-analysis map, and the attention map are input into the global branch, the ReHo index and the meta-analysis map are input into the local branch, and the output features of the global branch and the local branch are input into the fusion branch;

[0039] The ReHo index, the meta-analysis map, and the attention map are input and sequentially pass through the seven convolutional layers and the average pooling layer of the global branch; the ReHo index and the meta-analysis map are input and sequentially pass through the local patch extraction, the PSN network, and the average pooling layer of the local branch;

[0040] The output of the global branch and the output of the local branch sequentially pass through two subsequent FC layers; the ReLU function is used as the activation function for all activation functions; the meta-analysis map is incorporated into the network through the MAH network to guide the learning of the network.

[0041] Furthermore, the aforementioned deep learning system for brain disease discrimination guided by the hybrid of meta-analysis and individual attention further includes a model evaluation module for evaluating the model accuracy and specificity, and the model accuracy and specificity are evaluated as follows:

[0042] Accuracy calculation formula:

[0043] Specificity calculation formula:

[0044] Furthermore, when the aforementioned brain disease discrimination construction training module is training, the loss function is as follows:

[0045]

[0046] where the class label variable y iCharacterizing the clinical status attribution of the subject: y i = 0 corresponds to an individual in the healthy control group, y i = 1 indicates a patient diagnosed with a brain disease. The predicted confidence p(y i ) is the conditional probability estimator that the sample is correctly classified into its true label.

[0047] Compared with the prior art, the beneficial technical effects of the present invention adopting the above technical solutions are as follows:

[0048] (1) The present invention proposes a deep learning system and device guided by a mixture of meta-analysis and individual attention, providing a scientific basis for the clinical diagnosis of brain disease patients through objective imaging markers. This system operates completely automatically without manual intervention, thus effectively eliminating the deviation caused by subjective judgment in traditional diagnosis.

[0049] (2) The present invention designs an innovative solution that combines deep learning with domain knowledge in meta-analysis: the MAH network. In the face of the challenges of small sample size and strong individual heterogeneity of brain disease samples, by constructing two branches, global and local, the meta-analysis is integrated into the model as external domain knowledge. This design significantly improves the performance of the model, enhances its accuracy in disease discrimination, and at the same time improves the robustness and generalization ability of the model. In addition, by means of the method of meta-analysis atlas, an effective complementarity is achieved between the domain knowledge of brain disease research and the data-driven deep learning model.

[0050] (3) The present invention designs an attention mechanism in the deep learning model: the STA network. We designed the STA network to automatically identify abnormal locations related to brain diseases from the whole-brain ReHo through a weakly supervised method. Each subject can generate a unique attention map through the STA network, which is fused into the deep learning model through the constructed global branch to provide individual-level guidance to complete the high-accuracy discrimination of brain diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is a schematic diagram of the model construction process of the present invention;

[0052] Figure 2 is the structural diagram of the STA network of the present invention;

[0053] Figure 3 is the structural diagram of the MAH network of the present invention; DETAILED DESCRIPTION OF THE INVENTION

[0054] In order to better understand the technical content of the present invention, specific embodiments are hereby given and described in conjunction with the accompanying drawings as follows.

[0055] In the present invention, aspects of the present invention are described with reference to the accompanying drawings, in which a number of illustrative embodiments are shown. Embodiments of the present invention are not limited to those described in the drawings. It should be understood that the present invention can be implemented by any one of the various concepts and embodiments introduced above, as well as the concepts and embodiments described in detail below, because the concepts and embodiments disclosed in the present invention are not limited to any embodiment. In addition, some aspects disclosed in the present invention can be used alone, or in any suitable combination with other aspects disclosed in the present invention.

[0056] As Figure 1 shown, the present invention provides a deep learning system for brain disease discrimination based on a mixture of meta-analysis and individualized attention guidance, including:

[0057] ReHo feature extraction module, which obtains the ReHo index based on the local consistency of magnetic resonance imaging from brain images;

[0058] Meta-analysis atlas construction module, which is used to compare the significantly different voxel functions and structures and spatially convergent brain regions between patients with depression and healthy subjects, and obtain the neuroimaging feature atlas generated by meta-analysis;

[0059] Meta-analysis atlas integration and analysis module, which is used to establish a threshold screening mechanism and a spatial topology feature enhancement strategy for the neuroimaging feature atlas generated by meta-analysis, and obtain the standardized meta-analysis atlas M;

[0060] Brain disease discrimination construction and training module, including STA network and MAH network. The STA network is used to generate an attention map for each input brain image, and directly uses the classification label as the weak supervision localization guidance. The MAH network is used to fuse the meta-analysis atlas M, the ReHo index, and the attention map, and output the brain disease discrimination result.

[0061]

[0062] As a preferred embodiment of the present invention, the ReHo feature extraction module obtains the ReHo index based on the local consistency of magnetic resonance imaging from brain images. All subjects were scanned using a Siemens 3T MRI scanner (Siemens Healthcare). The resting-state fMRI was processed using the DPARSF toolbox based on SPM. The first six time points were discarded for magnetization equilibrium; motion alignment was performed to correct the images; the data was normalized into the MNI standard space; the images were band-pass filtered at 0.01 - 0.08 Hz, and the interference signals related to the global signal, cerebrospinal fluid, head motion (Friston 24 parameters), and white matter were regressed out. On this basis, the Kendal coefficient of concordance algorithm was used to perform neighborhood analysis on each voxel unit. By quantifying the degree of signal fluctuation coordination with the adjacent 26 voxels in the time dimension, the ReHo feature reflecting local neural activity synchronization was finally constructed. In this embodiment, 360 subjects (179 patients with brain diseases and 181 healthy controls) were used to establish a binary classification dataset by class label annotation. For the local consistency (ReHo) feature sequence, the sliding time window technique was used for time series segmentation (parameter settings: step size 6 time nodes, window length 30 nodes). By the sliding window strategy, time series feature segments were generated, achieving an enhancement effect with a data scale expansion coefficient of 18, effectively improving the stability of model training and the feature representation ability;

[0063] The specific calculation method is as follows:

[0064]

[0065] Among them, the functional magnetic resonance imaging (fMRI) time series contains N time nodes (N is the total number of scanning phases). For each specific time point i, the signal intensities of all voxels in the whole brain at this moment are sorted in ascending order through statistical methods to generate the sorting variable R. i . On this basis, represents the arithmetic mean of the R i sorting values corresponding to all time points, and the value of K is fixed at 27, corresponding to the number of adjacent voxels in the three-dimensional space neighborhood analysis (a cubic structure composed of the central voxel and its surrounding 26 voxels).

[0066] The sorting variable R of all voxels at time point i i is:

[0067]

[0068] Among them, the sorting of the signal value collected by voxel j at time node i in the time series is r i,j ;

[0069] The arithmetic mean of the R i sorting values corresponding to all time points is:

[0070]

[0071] As a preferred embodiment of the present invention, the meta-analysis atlas construction module constructs a meta-analysis atlas of the proposed hybrid deep learning framework, which includes 97 depression mechanism studies and 2,928 patients. We conducted searches on PubMed, Google Scholar, and BrainMap to identify neuroimaging experiments of major depressive disorder reporting gray matter atrophy, increased resting-state function, or decreased resting-state function compared to healthy controls. Using various combinations of search terms major depressive disorder, major depressive disorder, depressive disorder, unipolar depression, VBM, gray matter, regional cerebral blood flow, PET, SPECT, arterial spin labeling, regional homogeneity, ALFF / fALFF, glucose metabolism, brain activity, and resting state, VBM studies of regional cerebral blood flow and resting-state VBP studies, regional homogeneity, amplitude of low-frequency fluctuations (ALFF), and fractional ALFF (fALFF), as well as regional glucose metabolism were identified. The literature search was completed in 2024, determining an overall dataset of 102 independent publications reporting 169 experiments, where 5 publications included studies of structural and functional changes, for a total of 108 studies. Activation likelihood estimation (ALE) was used to generate a conjunction map of activation maps and test for spatial convergence above chance through various available threshold options. The method selected was cluster inference, which generates a simulated dataset with randomly distributed foci based on the characteristics of the input dataset to test the null hypothesis. In the screening of 108 previous studies, brain regions with significant differences and spatial convergence in voxel function and structure between patients with depression and healthy subjects were identified to obtain the meta-analysis atlas;

[0072] As a preferred embodiment of the present invention, for the neuroimaging feature atlas generated by the meta-analysis, we designed a preprocessing optimization process adapted to the deep learning framework. By establishing a threshold screening mechanism and a spatial topological feature enhancement strategy, the weight expression of disease-related brain regions was emphasized, thereby improving the multi-dimensional performance (classification accuracy, cross-dataset generalization ability) of the model. Based on the atlas-model co-optimization principle, dynamic weight d weight quantitative calculation, which is defined as:

[0073] d weight = log(PR(x))

[0074] Among them, the meta-analysis atlas is composed of discrete units in a three-dimensional space coordinate system. Each spatial unit x corresponds to a bibliometric parameter PR(x), which quantitatively characterizes the attention index of this brain region in historical research (the value range is 0-1, and PR(x)=0 means that this coordinate unit has not been reported in existing literature).

[0075] The mathematical representation of the meta-analysis atlas is defined as:

[0076]

[0077] Among them, d is the Euclidean distance from the center of the spatial unit to the activation peak point, and σ is the spatial attenuation coefficient of the Gaussian kernel function;

[0078] Furthermore, we construct an optimization index M for the spatial unit n , and its expression is:

[0079]

[0080] By introducing an empirical threshold to implement feature screening (units with M n <0.3 are regarded as low-confidence signals and removed), finally an optimized weight parameter M is generated, and the value of M is defined as:

[0081]

[0082] Based on this, a standardized meta-analysis atlas M is constructed as the input of the deep learning model.

[0083] As a preferred embodiment of the present invention, a brain disease discrimination construction and training module uses the ReHo feature extraction module's local consistency (ReHo) index based on magnetic resonance imaging and the main intention spectrum and meta-analysis atlas M as data carriers to construct an STA network and a MAH network for capturing discriminative knowledge guided by mixing and automatically diagnosing brain diseases. The deep learning brain disease discrimination model guided by the mixture of meta-analysis and individual attention is as Figure 1 shown. The discrimination model is divided into two stages: First, an STA network is designed to generate an attention atlas for each input brain image, and directly uses classification labels (such as MDD / HC) as weak supervision location guidance. The meta-analysis atlas and the attention atlas both have the same spatial resolution as the input ReHo, and each element represents the discriminative ability of the corresponding voxel for disease diagnosis. Guided by the meta-analysis atlas and the attention atlas, a MAH network is constructed to fuse the meta-analysis atlas and the attention atlas for brain disease discrimination;

[0084] Figure 2Shows the architecture of the STA network of the present invention, which consists of a convolutional (Conv) layer, a linear projection layer, two STF modules, an average pooling layer, and a classification (i.e., fully connected FC) layer. The convolutional layer uses a 2×2×2 kernel with a stride of 2; the linear projection layer uses a LayerNorm layer; the STF module consists of a LayerNorm layer, an MSA, and an MLP. The activation layer in the network uses the ReLU activation function. On the feature map generated by the last STF module, average pooling is performed to generate a feature vector, which is further used in the last FC layer (with softmax activation) to predict the probability score of the input ReHo belonging to a specific class. It should be noted that after network training, for the task-oriented localization of the FC layer, the bias is eliminated. We backpropagate the learned weights of the classification layer to the top Conv feature map to generate a spatial attention map for the corresponding ReHo. The attention map highlights the discriminative brain regions highly relevant to the diagnostic task. Specifically, the feature maps are {F1, F2,..., F M}, where the size of each F m (m = 1,..., M) is (X / 4)×(Y / 4)×(Z / 4), X×Y×Z is the size of the input ReHo, and M is the number of channels. According to the FC weights learned for the c-th class, i.e., the corresponding attention map A c is defined as:

[0085]

[0086] The classification score s c , i.e., the variable before softmax and normalization, is defined as:

[0087]

[0088] Finally, the attention maps of different classes are aggregated through quantization as:

[0089] A(x,y,z) = ∑ c s c A c (x,y,z),

[0090] where F m (x,y,z) is the feature map output by STA, is the parameter matrix of the FC layer after STA, A c is the attention map of one channel, A is the final attention map, and s c is the weight of each channel attention map.

[0091] The meta-analysis atlas can provide group-level guidance for the differential localization of brain diseases as external domain knowledge information; as a preferred embodiment of the present invention, the attention atlas can provide individualized guidance for the differential localization of brain diseases. The present invention designs a MAH network to integrate the meta-analysis atlas and the attention atlas for brain disease discrimination. The MAH network consists of a global branch, a local branch, and a fusion branch, as Figure 3 shown.

[0092] The global branch attempts to capture topic-specific discriminative information globally. To this end, the meta-analysis atlas and the intermediate feature map of the STA network, i.e., the attention atlas, are fused into the global branch. Since they have encoded the group-level information integrated from the literature by the meta-analysis and the individual-level information automatically extracted from the whole-brain ReHo for the diagnostic task. As the input to the global branch, the meta-analysis atlas and the attention atlas are spatially weighted across channels by voxel multiplication to amplify the influence of the input features on the potential information brain regions. After voxel local enhancement, they are further processed by seven convolutional layers to learn more discriminative feature maps. The third convolutional layer uses a 1×1×1 kernel, and the remaining convolutional layers all use a 3×3×3 kernel, followed by a normalization layer BN and a ReLU activation layer. Finally, an average pooling layer is added to generate the encoded topic-specific discriminative information.

[0093] The local branch extracts local discriminative information at the patch level. The local branch is built on top of the meta-analysis atlas positions, which are consistent in the ReHo space and are a stable complement to the global branch. The structure of the local branch is a hierarchical network that sequentially combines different patches to provide features for the diagnostic task. The input is multiple 3D image patches extracted from the linearly aligned ReHo. Each patch corresponds to a sub-network PSN, and the PSN network consists of six convolutional layers and two max pooling layers, and all convolutional layers use a 3×3×3 kernel. The PSNs share the same architecture and weights to limit the number of learnable parameters. Spatially adjacent patches are grouped into specific brain regions, and the outputs of the corresponding PSNs are connected according to their spatial relationships to form the input for the subsequent network. An average pooling layer is further used to generate the final feature representation in the local branch, encoding the consistent discriminative information from the local patches. Instead of using patches widely distributed on the whole-brain image as the network input, the present invention selects patches through the meta-analysis atlas to select positions in a more efficient data-driven manner. According to the meta-analysis atlas in the linearly aligned image space, voxels with values higher than a predefined threshold are selected as patches.

[0094] As Figure 3As shown, the fusion branch structure further fuses the global discriminative features (i.e., the output of the global branch) and the local discriminative features (i.e., the output of the local branch) through two subsequent FC layers, and then passes through a ReLU activation layer to learn an overall hybrid feature representation with higher discriminative ability. Finally, after passing through the softmax classification layer, the output is used for diagnosis.

[0095] The loss function of the deep learning model guided by the meta-analysis and individualized attention mixture. In the supervised learning framework, the binary cross entropy loss function commonly used in binary classification tasks is adopted to quantify the deviation degree between the predicted probability and the true label. The mathematical expression of this loss function is defined as follows:

[0096]

[0097] where the class annotation variable y i represents the clinical status attribution of the subject: y i = 0 corresponds to individuals in the healthy control group, and y i = 1 indicates patients diagnosed with brain diseases. The predicted confidence p(y i ) is the conditional probability estimator that the sample is correctly classified into its true label.

[0098] In the present invention, the training of the deep learning model guided by the meta-analysis and individualized attention mixture. First, a series of preprocessing steps (step (1)) are performed on the fMRI data used for model training to ensure data quality; then the construction and integrated analysis of the meta-analysis atlas are carried out; and then the processed ReHo index, classification label, and meta-analysis atlas are used as inputs and imported into the brain disease discrimination model. In this process, the STA network uses the whole-brain ReHo (size: 61×73×61) as the input, is trained for weakly supervised discriminative localization, and uses the subject's class label as the labeled true value. The MAH network uses the whole-brain ReHo, meta-analysis atlas, and attention atlas as inputs, and the meta-analysis atlas is integrated into the model through the local branch and the global branch to guide the training of the model. Through this training process, the parameters of the convolutional neural network are finally optimized to construct a deep learning system based on the meta-analysis and individualized attention mixture and its supporting device.

[0099] The testing of the deep learning model guided by the meta-analysis and individualized attention mixture. In the model testing stage, the above-mentioned preprocessing steps and the construction and integrated analysis of the meta-analysis atlas are also performed on the fMRI data, and the data is input into the trained brain disease discrimination model to discriminate the disease state.

[0100] Although the present invention has been described above with reference to preferred embodiments, it is not intended to limit the present invention. Those of ordinary skill in the art to which the present invention pertains can make various modifications and refinements without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A deep learning system for brain disease discrimination based on a hybrid guidance of meta-analysis and individualized attention, characterized in that Including: The ReHo feature extraction module, which obtains the ReHo index based on the local consistency of magnetic resonance imaging from brain images; The meta-analysis atlas construction module, which is used to compare the significantly different voxel functions and structures and spatially convergent brain regions between patients with depression and healthy subjects, and obtains the neuroimaging feature atlas generated by meta-analysis; The meta-analysis atlas integration and analysis module, which is used to establish a threshold screening mechanism and a spatial topology feature enhancement strategy for the neuroimaging feature atlas generated by meta-analysis, and obtains the standardized meta-analysis atlas M; The brain disease discrimination construction and training module, including the STA network and the MAH network. The STA network is used to generate an attention map for each input brain image, and directly uses the classification label as the weakly supervised localization guidance. The MAH network is used to fuse the meta-analysis atlas M, the ReHo index, and the attention map, and outputs the brain disease discrimination result.

2. The deep learning system for brain disease discrimination based on the hybrid guidance of meta-analysis and individualized attention according to claim 1, characterized in that, The ReHo index is calculated as follows: Among them, the functional magnetic resonance imaging (fMRI) time series contains N time nodes, where N is the total number of scanning phases. For each specific time point i, the signal intensities of all voxels in the whole brain at this moment are sorted in ascending order by statistical methods to generate a sorting variable R i , denotes R corresponding to all time points i is the arithmetic mean of the sorting values, and K is a fixed value corresponding to the number of adjacent voxels in the three-dimensional spatial neighborhood analysis.

3. The deep learning system for discriminating brain diseases based on the hybrid guidance of meta-analysis and individualized attention according to claim 2, wherein The sorting variable R of all voxels at time i i is as follows: Among them, the sorting of the signal value collected by voxel j at time node i in the time series is r i,j .

4. The deep learning system for discriminating brain diseases based on the hybrid guidance of meta-analysis and individualized attention according to claim 3, wherein The R corresponding to all time points i The arithmetic mean of the sorting values is as follows:

5. The deep learning system for brain disease discrimination based on the hybrid guidance of meta-analysis and individualized attention according to claim 1, characterized in that, The meta-analysis atlas integration and analysis module is configured to perform the following actions: S101. Perform dynamic weight d quantization calculation on each voxel unit, as follows: weight The following formula: d weight = log(PR(x)); Among them, each spatial unit x corresponds to the bibliometric parameter PR(x), which quantitatively represents the attention index of this brain region in historical studies, and the value range is 0-1. PR(x)=0 means that this coordinate unit has not been reported in existing literature; S102. The mathematical representation of the meta-analysis atlas is defined as: Among them, d is the Euclidean distance from the center of the spatial unit to the peak activation point, and σ is the spatial decay coefficient of the Gaussian kernel function; S103. Construct the spatial unit optimization index M n , as shown in the following formula: S104. Introduce an empirical threshold to implement feature screening. The units smaller than the empirical threshold are regarded as low-confidence signals and eliminated, and finally the optimized weight parameter M is generated: Based on this, a standardized meta-analysis atlas M is constructed as the input of the deep learning model.

6. The deep learning system for brain disease discrimination based on the hybrid guidance of meta-analysis and individualized attention according to claim 1, wherein The STA network includes a sequentially connected convolutional layer, a linear projection layer, a first STF module, a second STF module, a third STF module, an average pooling layer, and a classification layer; the STF module includes a LayerNorm layer, an MSA layer, and an MLP layer. The convolutional layer serves as the input end of the STA network, receives the ReHo index, and calculates the attention map as follows: Classification score s c : Aggregation of the quantified attention maps of different categories: A(x,y,z) = ∑ c s c A c (x,y,z), Among them, F m (x, y, z) is the feature map output by the STA, is the parameter matrix of the FC layer after the STA, A c is the attention map of one channel, A is the final attention map, s c is the weight of the attention map of each channel.

7. The deep learning system for brain disease discrimination based on the hybrid guidance of meta-analysis and individualized attention according to claim 3, wherein Taking the ReHo index, the meta-analysis atlas, and the attention map generated by the STA network as the input, and the brain disease discrimination result as the output, construct the MAH network; The ReHo index, the meta-analysis atlas, and the attention map are input into the global branch, and the ReHo index and the meta-analysis atlas are input into the local branch. The output features of the global branch and the local branch are input into the fusion branch; The ReHo index, the meta-analysis atlas, and the attention map are input and sequentially pass through the seven convolutional layers and the average pooling layer of the global branch; the ReHo index and the meta-analysis atlas are input and sequentially pass through the patch extraction, the PSN network, and the average pooling layer of the local branch; The output of the global branch and the output of the local branch sequentially pass through two subsequent FC layers; the ReLU function is used as the activation function for all activation functions; the meta-analysis atlas is integrated into the network through the MAH network to guide the learning of the network.

8. The deep learning system for brain disease discrimination based on the hybrid guidance of meta-analysis and individualized attention according to claim 1, characterized in that It also includes a model evaluation module, which is used to evaluate the accuracy and specificity of the model. The evaluation of the accuracy and specificity of the model is as follows: Accuracy calculation formula: Specificity calculation formula:

9. For the deep learning system for brain disease discrimination based on the hybrid guidance of meta-analysis and individualized attention according to claim 1, when the brain disease discrimination construction training module is training, the loss function is used as follows: Among them, Categorical annotation variable y i Characterize the clinical status attribution of the subject: y i = 0 corresponds to an individual in the healthy control group, y i = 1 indicates a patient diagnosed with a brain disease, and the predicted confidence p(y i ) is the conditional probability estimator that the sample is correctly classified into its true label.

Images