Core brain region screening method and system

Abstract

The invention belongs to the technical field of brain region function analysis, and particularly relates to a core brain region screening method and system. The core brain region screening method comprises the following steps: step 1, data acquisition and preprocessing; step 2, feature map construction and generation-discrimination network training; step 3, candidate sample generation, post-processing and quality evaluation; and 4, performing multi-algorithm feature consensus screening and stability evaluation. According to the core brain region screening method, firstly, class condition amplification with graph constraints is used, multi-index quality gating is assisted, then the reproducibility of feature selection is improved through multi-algorithm consensus screening and stability measurement, and finally, a stable and auditing judgment result is output in combination with multi-model integration. Therefore, the brain region screening performance and the generalization stability are remarkably improved under the scenes of scarcity of samples, class imbalance and high-dimensional structured data.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of brain region function analysis, and particularly relates to a core brain region screening method and system. BACKGROUND

[0002] Disease is a complex occurrence process. For example, when studying drug addiction, researchers often construct animal models such as drug self-administration and conditional place preference to reproduce the occurrence of addiction in experimental animals in order to study the specific mechanism of drug addiction. When further refining the research direction such as brain region, researchers often use c-Fos, electroencephalogram and other data to show the activity of different brain regions, which is used as evidence for screening key brain regions related to drug addiction. However, such data often have characteristics such as "small sample, class imbalance and significant structure correlation", which directly lead to insufficient statistical power, increased risk of overfitting and limited cross-laboratory reproducibility, bringing challenges to reliable screening of key brain regions.

[0003] In addition, the existing screening methods often rely on manual selection, univariate statistical test or single algorithm feature selection, and often cannot depict the correlation structure between brain regions and global network constraints. At the same time, common data augmentation methods such as simple oversampling or generative models without considering graph structure can easily destroy the original distribution and correlation structure, and the setting of the augmentation rate lacks objective quality gating and controlled injection mechanism, which is easy to introduce distribution drift and "noise amplification". On the other hand, the ranking results of different feature selection algorithms differ greatly, and there is a lack of "multi-algorithm consensus" and stability measurement for small sample and unbalanced data. Related studies have also shown that in the intelligent discrimination of neural data, the effectiveness of machine learning is highly dependent on the construction and quality control of the feature set, and the existing feature set and process often cannot meet the dual requirements of precision and robustness. SUMMARY

[0004] The purpose of the present application is to provide a core brain region screening method and system.

[0005] The first aspect of the present application provides a core brain region screening method, comprising: Step 1, data acquisition and preprocessing; Step 2, feature graph construction and generation-discriminant network training; Step 3, candidate sample generation, post-processing and quality evaluation; Step 4, multi-algorithm feature consensus screening and stability evaluation.

[0006] In an embodiment of the present application, the data acquisition in step 1 comprises: 8-10 weeks old male rats in the treatment group at the end of the addiction formation stage and the control group were taken as "samples", and the c-Fos neuronal activity data obtained by immunofluorescence staining of the tissues of different brain regions of each rat were taken as "features", and a "sample-feature" data matrix was constructed.

[0007] In an embodiment of the present application, the data preprocessing in step 1 includes: Reading the matrix data, and performing non-negativity conversion on the feature values in the matrix; Performing Log conversion on the non-negative feature values; Performing Z-score standardization processing on the feature values, calculating the mean and standard deviation of each feature value, and converting all feature values into a standardized feature matrix with a mean of 0 and a standard deviation of 1; Based on the standardized feature matrix and the corresponding grouping information label, a stratified random sampling method is used to divide the training set and the test set, and the division ratio is 7:3. After division, the training set is used for subsequent model training, and the test set is used for final model generalization performance evaluation.

[0008] In an embodiment of the present application, the method for constructing a feature graph in step 2 includes: Performing rank transformation and center standardization on each feature, and clipping the correlation coefficient to [-1, 1]; Based on the robust correlation, the correlation matrix is calculated based on the two features of each feature; Taking the absolute value of each correlation value to obtain a similarity matrix, and performing top-k selection on each row of the similarity matrix to retain the strongest k connections, and clearing the remaining elements, and then performing symmetrization to obtain an undirected adjacency A, and then performing degree normalization based on the degree matrix D to obtain a feature graph .

[0009] In an embodiment of the present application, the method for training a discriminative network in step 2 includes: A WGAN generative adversarial model GraphMix-WGAN containing a GraphMix layer is constructed on the feature graph; wherein the GraphMix performs graph smoothing mixing on the input feature matrix H, and uses the residual formula mix = (1-m)H + m H to retain individual feature information while injecting neighborhood structure, that is, a feature graph ; wherein H is a "sample-feature" data matrix of the treatment group and the control group; m is a mixing coefficient of GraphMix, used to control the injection strength of neighborhood information; H mix is the output feature matrix of GraphMix, which has the same size and shape as H; The generator end adopts Softplus as the output activation to obtain non-negative and gradient-continuous output; the structure is "random noise input→two-layer fully connected backbone→one layer of GraphMix to inject graph structure→fully connected mapping to feature dimension→Softplus output"; The discriminator structure is "real matrix input→one layer of GraphMix to align graph prior→multiple layers of fully connected blocks with spectral normalization→real or fake discrimination score"; In the training process, the Wasserstein norm is used, and two types of stabilizing terms are superimposed: WGAN-GP gradient penalty and R1 regularization.

[0010] In an embodiment of the present application, the method of candidate sample generation in step 3 includes: Independent sampling of the generator in the respective training period "best" checkpoint of Saline and METH groups: when loading the checkpoint, the EMA shadow weight is restored first to improve the inference stability, and the feature map and standardization scale saved during the training stage are also retrieved to ensure that the sampling and training distributions are strictly aligned; then generate Gaussian noise input by batch, get the "standardized logarithmic domain" output by forward through the generator and the feature map, restore it to the logarithmic domain using the scale, and then return to the original count domain using the expm1 function, and do the minimum value truncation to avoid non-positive numbers, thus obtaining the candidate sample.

[0011] In an embodiment of the present application, the method of post-processing in step 3 includes: In the logarithmic domain, the relevant structure of the synthetic sample is "moved" to be consistent with the real sample by the rank-based U transformation and Gaussian copula; The piecewise monotonic smoothing inverse CDF is used to restore the marginal distribution of each dimension; The quantile matching is performed again to correct the residual marginal difference; The histogram locking based on 32 barrels is implemented to make the frequency band proportion of each feature dimension strictly consistent with the real data, and after completion, it is mapped back to the original count domain and non-negative truncated.

[0012] In an embodiment of the present application, the method of quality evaluation in step 3 includes: For each augmented candidate, the maximum mean difference MMD and the edge JS divergence of the respective training measured set are calculated, and the slice Wasserstein distance of 64 random projections is used to supplement the shape difference measure; wherein When the number of real and synthetic samples is inconsistent, the SWD is calculated after random downsampling to the same size to avoid sample number bias, and the smaller the three indicators are, the closer to the real distribution they are, and then the comprehensive quality score Q is obtained by weighting and aggregating the respective training sample number by category (group). The best amplification ratio is selected by using a "rank and" criterion for evaluating each preset amplification ratio set.

[0013] In an embodiment of the present application, the method of multi-algorithm feature consensus screening and stability evaluation in step 4 comprises: On the basis of the amplified data set, multi-source feature screening is performed in parallel, and the importance of the graph structure is fused to form the final Top-N: The abnormal samples are removed and standardized for the input amplified data; At least three of ANOVA, LASSO, random forest, recursive feature elimination RFE, and mutual information MI are selected to calculate the importance ranking of each brain region in parallel; The rankings of each method are unified to the same index, the ranking average of all brain regions is obtained to obtain the "consensus ranking", and the ranking is mapped to a 0-1 normalized score for subsequent weighting; The relevant kNN graph is constructed and the GNN-lite is trained; The final ranking is obtained by linear fusion.

[0014] Correspondingly, the second aspect of the present application provides a core brain region screening system, comprising: A data acquisition and preprocessing module for data acquisition and preprocessing; A feature map construction module for constructing a feature map; A network training module for generating a discriminant network training; A candidate sample processing module for generating, post-processing and quality evaluation of candidate samples; A screening and evaluation module for multi-algorithm feature consensus screening and stability evaluation.

[0015] The core brain region screening method of the present application first uses a class condition amplification with graph constraints and is supplemented by multi-index quality gating, then uses multi-algorithm consensus screening and stability measurement to improve the reproducibility of feature selection, and finally combines multi-model integration to output a stable and auditable discriminant result, thereby significantly improving the brain region screening performance and generalization stability in the scene of sample scarcity, class imbalance and high-dimensional structured data.

[0016] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent to those skilled in the art from the description, or can be learned by practice of the present application. The objects and other advantages of the present application will be realized and achieved by the structures particularly pointed out in the specification, claims, and drawings.

[0017] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the following preferred embodiments are specifically described below, and the accompanying drawings are described as follows. BRIEF DESCRIPTION OF DRAWINGS

[0018] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a flowchart of a core brain region screening method according to an embodiment of the present invention; Figure 2 This is a flowchart of the quality assessment of candidate sample generation according to an embodiment of the present invention; Figure 3 This is a flowchart of a multi-algorithm feature consensus screening and stability evaluation according to an embodiment of the present invention; Figure 4 This is an example diagram of the ranking obtained by screening according to an embodiment of the present invention; Figure 5 This is a schematic diagram of a core brain region screening system according to an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] like Figure 1 As shown, one embodiment of the present invention provides a method for screening core brain regions, which includes: Step 1, Data Acquisition and Preprocessing; Step 2, Feature Map Construction and Generation—Discriminator Network Training; Step 3: Candidate sample generation, post-processing, and quality assessment; Step 4: Multi-algorithm feature consensus screening and stability evaluation.

[0022] Specifically, the data acquisition in step 1 includes: using 8-10 week old male rats in the late stage of addiction formation and control group as "samples", and using c-Fos neuron activity data obtained by immunofluorescence staining of different brain regions of each rat as "features", and constructing a "sample × feature" data matrix.

[0023] Data preprocessing includes: reading the matrix data, then performing a non-negativity transformation on the eigenvalues, converting all values ​​less than 0 to 0 to ensure non-negativity. Next, a Log transformation is performed on the non-negative eigenvalues ​​to narrow their range and make them closer to a normal distribution. Finally, Z-score standardization is applied to the eigenvalues, calculating the mean and standard deviation of each eigenvalue, converting all eigenvalues ​​into a standardized feature matrix with a mean of 0 and a standard deviation of 1. Furthermore, based on the standardized feature matrix and corresponding group labels, stratified random sampling is used to divide the training and test sets in a 7:3 ratio. The training set is used for subsequent model training, while the test set is only used for final model generalization performance evaluation. The input unprocessed matrix instance is shown in the figure below (only partial data is shown). Different brain regions represent features, columns represent different samples, and numerical values ​​represent the activity level of the brain region. Saline (saline group, i.e., control group) and METH (metamine addiction group, i.e., treatment group) are the group labels.

[0024] Furthermore, step 2 specifically includes: Feature map construction: To reduce the impact of outliers, the rank transformation and center standardization of each feature are first performed, and the correlation coefficient is clipped to [-1,1]. Then, the correlation matrix is ​​obtained by calculating each feature pairwise based on the robust correlation pair matrix. Then, the absolute value of each correlation value is taken to obtain the similarity matrix. In each row of the similarity matrix, top-k neighbor selection is performed to retain the k strongest connections, and the remaining elements are cleared to zero. Then, symmetry is performed to obtain the undirected adjacency A. Finally, degree normalization is performed based on the degree matrix D to obtain the feature map. This characterizes the robust structural similarity between brain regions. In this embodiment, the default value of k is 5, which can be adjusted between 3 and 8 depending on the number of features and the noise level.

[0025] Generative-discriminative network training: A generator and a discriminator, each containing a GraphMix layer, are constructed on the feature map, forming a WGAN generative adversarial model with GraphMix layers, GraphMix-WGAN. GraphMix performs graph smoothing mixing on the input feature matrix H, preferentially using H... mix =(1-m)H + m H The residual formula preserves individual feature information while injecting neighborhood structure, i.e., feature map. Where H is the feature matrix, i.e., the "sample × feature" data matrix of the Saline and METH groups; m is the mixing coefficient of GraphMix, used to control the intensity of neighborhood information injection; H mixThis is the output feature matrix of GraphMix, with the same size and shape as H. The recommended range for the mixing coefficient m is 0.02 - 0.08; in this embodiment, m = 0.05 is used. Optionally, the feature map... It can be a structure trained based on all the data. During training, the matrix H can be trained separately for different groups, i.e., one H for the Saline group and one H for the METH group.

[0026] The generator employs Softplus as the output activation to obtain a non-negative and gradient-continuous output. Its structure is: "random noise input → two fully connected backbone layers → one GraphMix injection graph structure layer → fully connected mapping to the feature dimension → Softplus output". Softplus is used to obtain a smooth and non-negative output, while applying an exponential moving average (EMA) to the generator parameters to enhance stability, where the decay coefficient of the EMA is α = 0.999. The discriminator structure is "real matrix input → one layer of GraphMix aligned graph prior → multiple layers of fully connected blocks with spectral normalization → true / false discrimination score". Here, spectral normalization is used to constrain Lipschitz behavior to suppress gradient explosion. In addition, in order to improve the adversarial coverage between the generator and the discriminator, a strategy of multiple discriminators is adopted. In this study, the number of discriminators is 3. During training, the Wasserstein paradigm is employed with two stabilizing terms: WGAN-GP gradient penalty and R1 regularization. The former, with a coefficient λ_gp=10, primarily constrains the discriminator's gradient on interpolated samples, while the latter, with a coefficient γ_r1=1, smooths the discrimination of real data. The training update ratio is 3:1 (discriminator:generator) to first stabilize the discrimination boundary before driving the generator to approximate the real distribution. Dynamic regularization weights are enabled: gradually increasing weights in the first half of training and maintaining or slightly decreasing them in the second half, achieving a better balance between "adversarial" and "distribution fit." Validation set performance and generalization metrics are used as the basis for early stopping and model selection. Throughout the training process, neighborhood information is entirely provided by the constructed feature maps and remains static during training. Furthermore, in this example, the feature maps... The construction is based on the full feature matrix, while during GraphMix-WGAN training, two sets of feature matrices are trained separately based on the group labels to maintain the differences between groups.

[0027] Further, see Figure 2 Step 3 specifically includes: Candidate sample generation: For the Saline and METH groups, the generators in their respective "best" checkpoints during training are sampled independently: when loading checkpoints, the EMA shadow weights are restored first to improve inference stability, and the feature maps and standardized scales saved during the training phase are retrieved at the same time to ensure that the sampling is strictly aligned with the training distribution; then Gaussian noise input is generated in batches, and the "standardized logarithmic domain" output is obtained by forward passing through the generator and feature maps. The scale is used to restore it to the logarithmic domain, and then the expm1 function is used to return to the original count domain. The minimum value is truncated to avoid non-positive numbers, and the candidate samples are obtained.

[0028] Post-processing is performed on the candidate samples: First, in the logarithmic domain, the relevant structure of the synthetic sample is "transferred" to be consistent with the real sample through rank-based U-transform and Gaussian copula. Then, the edge distribution of each dimension is recovered using piecewise monotonic smooth inverse CDF. Next, quantile matching is performed to correct residual edge differences. Finally, histogram locking based on 32 buckets is implemented to ensure that the frequency band proportion of each feature dimension is strictly consistent with the real data. After completion, the original count domain is mapped back from the logarithmic domain and non-negative truncation is performed. The quality assessment phase is completed in the logarithmic domain: For each amplification candidate, the maximum mean difference (MMD) and marginal JS divergence with their respective training and measured sets are calculated, and the shape difference measure is supplemented by the Wasserstein distance (SWD) of 64 randomly projected slices. It is worth noting that when the number of real and synthetic samples is inconsistent, the samples are first randomly downsampled to the same size before calculating the SWD to avoid sample number bias. The smaller the three indicators are, the closer they are to the real distribution. Then, they are aggregated by category (group) with their respective training sample numbers to obtain the comprehensive quality score Q.

[0029] The pre-defined amplification ratio set was evaluated one by one, using a "ranking sum" criterion. This involves summing the MMD, JS, and SWD values ​​in ascending order, with MMD as the priority criterion in case of ties, to select the optimal amplification ratio. In this example, six amplification ratios of 2, 4, 6, 8, 10, and 12 were set. Empirical results showed that ×8 was the most stable, so this ratio was ultimately adopted. Data injection was performed based on the class-adaptive injection rate and neighborhood density constraints: the number of synthetic samples injected for each class = the original training number for that class × (optimal amplification ratio - 1), resulting in a new training set size of "original × optimal amplification ratio," while the test set remained unchanged.

[0030] The specific details of the overall quality score Q under different injection ratios in this embodiment are shown in the table below.

[0031] Further, see Figure 3 The methods for multi-algorithm feature consensus screening and stability evaluation in step 4 include: See Figure 3 Based on the augmented dataset, multi-source feature filtering is performed in parallel, and graph structure importance is fused to form the final Top-N: First, abnormal samples were removed and standardized from the input amplified data, and then ANOVA and other standardization methods were applied. At least three methods—LAASSO, Random Forest, Recursive Feature Emission (RFE), and Mutual Information Mining (MI)—are selected to compute the importance ranking of each brain region in parallel. Then, the rankings from each method are unified to the same index, and the average ranking across all brain regions is calculated to obtain a "consensus ranking." This ranking is then mapped to a 0–1 normalized score for subsequent weighting. Specifically, the parameters for LASSO are: solver=saga, max_iter=10000, class_weight=balanced, C=1.0; the parameters for Random Forest are: class_weight=balanced, n_estimators=500; the parameters for RFE are: base learner is linear SVM: C=1.0, dual=False, max_iter=5000; the default number of features retained is approximately √(current number of features); and the parameters for MI are: n_neighbors=3. Based on the above five methods, a related kNN graph is further constructed and GNN-lite is trained: the absolute value of feature correlation is calculated as the similarity using the training data, k=8 nearest neighbors are selected for each feature and symmetricized, self-loops are added and degree is normalized to obtain the graph; logistic regression is trained on the "graph smoothed features" on the hierarchical k-fold, the absolute value of the model coefficients of each fold is averaged as the graph perception importance and the full ranking table is calculated, where k≤5 and automatically converges with the number of minority class samples; finally, the final ranking is obtained by linear fusion: the consensus score and the GNN-lite normalized importance are read, and Consensus*(1−w) + GNN*w is fused according to the weight coefficient w=1 / 6. The above five methods account for 5 / 6, and GNN accounts for 1 / 6. The brain region rankings obtained by each algorithm are normalized and the consensus ranking is obtained by the average ranking of Borda counts. Based on this, the final Top-N list is generated, and the best brain region ranking can be obtained by screening to select the most important brain region. Figure 4 This is a ranking example obtained by filtering the sample data in step 1 in this embodiment. The vertical axis represents the ranking of brain regions with "features" obtained through consensus filtering using multiple methods. Only the top 20 brain regions are shown in the figure. The horizontal axis represents the consensus importance score obtained after normalization. The closer the score is to 1, the higher its importance.

[0032] See Figure 5 Based on the above embodiments, an embodiment of the present invention also provides a core brain region screening system, comprising: The data acquisition and preprocessing module is used for data acquisition and preprocessing. The feature map construction module is used to construct feature maps; The network training module is used to generate and discriminate network training. The candidate sample processing module is used for candidate sample generation, post-processing, and quality assessment. The screening and evaluation module is used for screening and evaluating the stability of multi-algorithm feature consensus.

[0033] The working methods of each module of the core brain region screening system in this embodiment can be found as described above, and will not be repeated here.

[0034] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative; for example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0035] In addition, the functional modules in the various embodiments of the present invention can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0036] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0037] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for screening core brain regions, characterized in that, include: Step 1, Data Acquisition and Preprocessing; Step 2, Feature Map Construction and Generation—Discriminator Network Training; Step 3: Candidate sample generation, post-processing, and quality assessment; Step 4: Multi-algorithm feature consensus screening and stability evaluation.

2. The core brain region screening method as described in claim 1, characterized in that, The data acquisition in step 1 includes: Male rats aged 8-10 weeks, in the late stage of addiction formation and in the control group, were used as "samples". The c-Fos neuronal activity data obtained by immunofluorescence staining of different brain regions of each rat were used as "features". A "sample × feature" data matrix was constructed.

3. The core brain region screening method as described in claim 2, characterized in that, The data preprocessing in step 1 includes: Read the matrix data and perform a nonnegativity transformation on the eigenvalues ​​in the matrix; Perform Log transformation on non-negative eigenvalues; Perform Z-score standardization on the eigenvalues, calculate the mean and standard deviation of each eigenvalue, and convert all eigenvalues ​​into a standardized feature matrix with a mean of 0 and a standard deviation of 1. Based on the standardized feature matrix and corresponding group information labels, a stratified random sampling method is used to divide the training set and the test set, with a division ratio of 7:

3. After the division, the training set is used for subsequent model training, and the test set is used for the final model generalization performance evaluation.

4. The core brain region screening method as described in claim 3, characterized in that, The method for constructing the feature map in step 2 includes: Perform rank transformation and center standardization on each feature to trim the correlation coefficient to [-1, 1]; The correlation matrix is ​​obtained by calculating the correlation pair matrix based on the features; The absolute value of each correlation value is taken to obtain a similarity matrix. Then, top-k neighbor selection is performed in each row of the similarity matrix to retain the k strongest connections, and the remaining elements are cleared to zero. Symmetry is then applied to obtain undirected adjacency A. Finally, degree normalization is performed based on the degree matrix D to obtain the feature map. .

5. The core brain region screening method as described in claim 4, characterized in that, The method for generating-discriminating network training in step 2 includes: A WGAN generative adversarial model, GraphMix – WGAN, containing GraphMix layers, is constructed on the feature maps. GraphMix performs graph smoothing mixing on the input feature matrix H, using H... mix =(1-m)H + m H The residual method preserves individual feature information while injecting neighborhood structure, i.e., feature map. Where H is the feature matrix, i.e., the "sample × feature" data matrix of the treatment group and the control group; m is the mixing coefficient of GraphMix, used to control the intensity of neighborhood information injection; H mix H is the output feature matrix of GraphMix, which has the same size and shape as H. The generator uses Softplus as the output activation to obtain a non-negative and gradient-continuous output; its structure is "random noise input → two fully connected backbones → one GraphMix injection graph structure → fully connected mapping to feature dimension → Softplus output"; The discriminator structure is "real matrix input → one layer of GraphMix aligned graph prior → multiple layers of fully connected blocks with spectral normalization → true / false discrimination score"; During training, the Wasserstein paradigm is used and two types of stabilizing terms are superimposed: WGAN-GP gradient penalty and R1 regularization.

6. The core brain region screening method as described in claim 5, characterized in that, The method for generating candidate samples in step 3 includes: For the Saline and METH groups, the generators in their respective "best" checkpoints during training are sampled independently: when loading checkpoints, the EMA shadow weights are restored first to improve inference stability, and the feature maps and standardized scales saved during the training phase are retrieved at the same time to ensure that the sampling is strictly aligned with the training distribution; then Gaussian noise input is generated in batches, and the "standardized logarithmic domain" output is obtained by forward passing through the generator and feature maps. The scale is used to restore it to the logarithmic domain, and then the expm1 function is used to return to the original count domain, and the minimum value is truncated to avoid non-positive numbers. At this point, candidate samples are obtained.

7. The core brain region screening method as described in claim 6, characterized in that, The post-processing method in step 3 includes: In the logarithmic domain, the relevant structure of the synthetic sample is "transferred" to be consistent with that of the real sample through rank-based U-transform and Gaussian copula. The edge distributions of each dimension are recovered using a piecewise, monotonic, smoothed inverse CDF. Perform another quantile matching to correct residual edge differences; Implement histogram locking based on 32 buckets to ensure that the frequency band proportion of each feature dimension is strictly consistent with the real data. After completion, map back from the logarithmic domain to the original count domain and perform non-negative truncation.

8. The core brain region screening method as described in claim 7, characterized in that, The quality assessment methods in step 3 include: For each amplification candidate, the maximum mean difference (MMD) and marginal JS divergence with their respective training and measured sets are calculated, and the shape difference metric is supplemented by the Wasserstein distance of slices from 64 random projections; where When the number of real and synthetic samples is inconsistent, the sample size is first randomly downsampled to the same size before calculating SWD to avoid sample number bias. The smaller the three indicators are, the closer they are to the real distribution. Then, the samples are aggregated by category (group) with their respective training sample numbers to obtain the comprehensive quality score Q. Each preset set of amplification rates is evaluated, and the optimal amplification rate is selected using the "ranking sum" criterion.

9. The core brain region screening method as described in claim 8, characterized in that, The methods for multi-algorithm feature consensus screening and stability evaluation in step 4 include: Based on the expanded dataset, multi-source feature filtering is performed in parallel, and graph structure importance is fused to form the final Top-N: Complete the abnormal sample removal and standardization of the input amplified data; At least three of the following methods were selected to calculate the importance ranking of each brain region in parallel: ANOVA, LASSO, random forest, recursive feature elimination (RFE), and mutual information (MI). The rankings of each method are unified to the same index, the rankings of all brain regions are averaged to obtain the "consensus ranking", and the rankings are mapped to 0-1 normalized scores for subsequent weighting. Construct relevant kNN graphs and train GNN-lite; The final ranking is obtained by linear fusion.

10. A core brain region screening system, characterized in that, include: The data acquisition and preprocessing module is used for data acquisition and preprocessing. The feature map construction module is used to construct feature maps; The network training module is used to generate and discriminate network training. The candidate sample processing module is used for candidate sample generation, post-processing, and quality assessment. The screening and evaluation module is used for screening and evaluating the stability of multi-algorithm feature consensus.

Images