Brain information detection method based on multi-atlas and combined feature selection

Abstract

The invention provides a brain information detection method based on multi-atlas and combined feature selection, and the method comprises the steps: carrying out the standardization processing of collected rs-fMRI data, and obtaining the preprocessing data; performing multi-template feature extraction on the preprocessed data to obtain an FC feature vector; performing three-layer progressive feature screening on the FC feature vector to obtain a core discriminative feature; optimizing and verifying a kernel function and a hyper-parameter of the classifier according to the core discriminative features to obtain a trained classification model; and carrying out key ROI and brain network detection on the core discriminative features through a classification model to obtain a detection result. According to the method, a multi-stage feature selection strategy is integrated, and a classification model with strong adaptability is adopted, so that the problems of high feature dimension, single screening, strong template dependence and the like in the prior art are solved, and the accuracy, the sensitivity, the specificity and the robustness of brain region detection are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of clinical medical technology, and in particular to a method for detecting brain information based on multi-map and combined feature selection. Background Technology

[0002] Mild traumatic brain injury (mTBI) is a common and prevalent type of brain injury caused by external impact, which can lead to long-term cognitive impairment and pose a serious challenge to public health. In recent years, resting-state functional magnetic resonance imaging (rs-fMRI) has emerged as a non-invasive and easy-to-operate technique, revealing brain network connectivity patterns by recording spontaneous brain activity in patients at rest. To convert fMRI signals into analyzable brain region features, the primary method relies on predefined brain atlases to construct functional connectivity (FC), including Automated Anatomical Labeling (AAL) based on anatomical structures. AAL is accurate and widely used, but its sensitivity to brain functional abnormalities is limited. Power-264 and Brainnetome-246 brain templates, based on functional construction, are used to capture brain network topological features. However, different brain templates employ different segmentation strategies, and single-template analysis can easily lead to result bias and generally suffers from the following problems: 1. Curse of Dimensionality: rs-fMRI data contains a large number of Regions of Interest (ROIs), and the functional connectivity features constructed can reach tens of thousands of dimensions. This can easily lead to model overfitting, reduce the reliability of diagnostic results, and significantly increase training difficulty and computational cost.

[0003] 2. Limited Feature Selection Strategies: Existing technologies often employ single feature selection methods, such as LASSO and T-tests, which make it difficult to simultaneously consider feature relevance, importance, and sparsity. This makes it difficult to effectively screen out highly discriminative features and comprehensively capture multi-network anomaly features of mTBI.

[0004] 3. Limitations of the applicability of brain templates: The brain regions and network structures extracted from different brain templates vary significantly. Existing technologies mostly rely on a single template for analysis, which can easily lead to biased results and lacks robustness for cross-template validation. This results in poor consistency of diagnostic results under different templates, making it difficult to guarantee the accuracy and stability of the results.

[0005] 4. Insufficient adaptability of classification models: Some methods use classifiers, such as KNN and Bayesian classifiers, which are difficult to capture the nonlinear relationships in high-dimensional FC features, and the classification accuracy needs to be improved. Summary of the Invention

[0006] The purpose of this invention is to provide a brain information detection method based on multi-map and combined feature selection, which improves the detection and classification accuracy of brain regions through multi-stage feature selection strategies and classification models.

[0007] To achieve the above objectives, the present invention provides the following solution: A brain information detection method based on multi-map and combined feature selection includes the following steps: The acquired rs-fMRI data were standardized to obtain preprocessed data; Based on the first and second templates, multi-template feature extraction is performed on the preprocessed data to obtain FC feature vectors; The FC feature vector is subjected to a three-level progressive feature selection to obtain the core discriminative features; The kernel function and hyperparameters of the classifier are optimized and validated based on the core discriminative features to obtain the trained classification model; The core discriminative features were analyzed using a classification model to detect key ROIs and brain networks, and the results were obtained. The results included node degree and distribution of overlapping brain regions.

[0008] Optionally, the acquired rs-fMRI data are standardized to obtain preprocessed data, including: Time-corrected data were obtained by performing time correction on the rs-fMRI data. The time-corrected data is subjected to head movement correction and data alignment to obtain aligned data; The aligned data is rigidly registered and segmented with the T1 structural image to obtain the registered data; The registered data is mapped to a standard template and then spatially smoothed and bandpass filtered to obtain preprocessed data.

[0009] Optionally, multi-template feature extraction is performed on the preprocessed data based on the first and second templates to obtain an FC feature vector, including: The preprocessed data is divided into regions of interest using the first template and the second template, respectively, to obtain the first time series and the second time series. Calculate the Pearson correlation coefficients for the first and second time series respectively, and construct the FC matrix based on the Pearson correlation coefficients; Extract and exclude the lower triangular elements of the diagonal of the FC matrix to obtain the FC eigenvectors.

[0010] Optionally, a three-layer progressive feature selection is performed on the FC feature vector to obtain core discriminative features, including: Mutual information filtering is performed on the FC feature vectors to obtain the first-level filtered features; Random forest feature importance filtering is applied to the primary selection features to obtain secondary selection features; LASSO screening was performed on the secondary screening features to obtain the core discriminative features.

[0011] Optionally, the kernel function and hyperparameters of the classifier are optimized and validated based on the core discriminative features to obtain the trained classification model, including: The training set was obtained by performing 20 rounds of 5-fold cross-validation on the core discriminative features; The optimal parameters for the classifier's regularization and RBF kernel were optimized using grid search. The classifier is trained on the training set based on the optimal parameters to obtain the initial model; The initial model is validated across templates using core discriminative features to obtain a classification model.

[0012] Optionally, the kernel function and hyperparameters of the classifier are optimized and validated based on the core discriminative features to obtain the trained classification model. The method also includes: evaluating the classification model using evaluation metrics, including accuracy, sensitivity, specificity, and area under the curve.

[0013] Optionally, key ROI and brain network detection are performed on the core discriminative features using a classification model to obtain detection results, including: The core discriminative features are functionally decomposed using a classification model to obtain functional connectivity pairs between the first and second templates; the functional connectivity pairs include the first and second brain regions. Traverse all function connection pairs to obtain the node degree; Based on standard spatial coordinates, coordinate mapping is performed on the first and second templates to obtain the distribution of overlapping brain regions.

[0014] Optionally, coordinate mapping is performed on the first and second templates based on standard spatial coordinates to obtain the distribution of overlapping brain regions, including: Obtain the three-dimensional center coordinates of the first brain region of the second template and the brain region coordinates of the second brain region of the second template; Determine whether the three-dimensional center coordinates of the first brain region of the second template and the three-dimensional center coordinates of the second brain region of the second template are within the range of the brain region coordinates of the first brain region and the second brain region of the first template, and summarize all the coordinates that pass the judgment to obtain the overlapping brain regions across templates. By mapping overlapping brain regions across templates to different core brain networks, the distribution of overlapping brain regions is obtained.

[0015] Optionally, the core brain networks include: the default mode network, the sensorimotor network, the subcortical network, the attention network, the visual network, and the cerebellar network.

[0016] According to specific embodiments provided by the present invention, the following technical effects are disclosed: The brain information detection method based on multi-map and combined feature selection provided by the present invention includes: standardizing the acquired rs-fMRI data to obtain preprocessed data; extracting multi-template features from the preprocessed data according to a first template and a second template to obtain FC feature vectors; performing a three-layer progressive feature selection on the FC feature vectors to obtain core discriminative features; optimizing and validating the kernel function and hyperparameters of the classifier based on the core discriminative features to obtain a trained classification model; and performing key ROI and brain network detection on the core discriminative features through the classification model to obtain detection results; the detection results include: node degree and distribution of overlapping brain regions. This method integrates a multi-stage feature selection strategy and adopts a highly adaptable classification model, solving the problems of high feature dimensionality, single selection, and strong template dependence in existing technologies, effectively improving the accuracy, sensitivity, specificity, and robustness of brain region detection. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the brain information detection method based on multi-map and combined feature selection of the present invention; Figure 2 This is a schematic diagram illustrating the working principle of brain information detection in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments 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, and 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.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] like Figure 1and Figure 2 As shown, this invention provides a brain information detection method based on multi-map and combined feature selection, comprising the following steps: Step 100: Standardize the acquired rs-fMRI data to obtain preprocessed data; Step 200: Perform multi-template feature extraction on the preprocessed data based on the first template and the second template to obtain the FC feature vector; Step 300: Perform a three-layer progressive feature filtering on the FC feature vector to obtain the core discriminative features; Step 400: Optimize and validate the kernel function and hyperparameters of the classifier based on the core discriminative features to obtain the trained classification model; Step 500: Use a classification model to perform key ROI and brain network detection on the core discriminative features to obtain the detection results; the detection results include: node degree and distribution of overlapping brain regions.

[0022] Preferably, the acquired rs-fMRI data is standardized to obtain preprocessed data, including: Time-corrected data were obtained by performing time correction on the rs-fMRI data. The time-corrected data is subjected to head movement correction and data alignment to obtain aligned data; The aligned data is rigidly registered and segmented with the T1 structural image to obtain the registered data; The registered data is mapped to a standard template and then spatially smoothed and bandpass filtered to obtain preprocessed data.

[0023] In the specific implementation process, step 100 uses MATLAB 2013a and SPM12 software for data preprocessing, and the specific steps are as follows: Temporal correction: Based on the temporal differences of slice scanning, the built-in algorithm of SPM12 is used to compensate for time offset; Head movement correction: The translation and rotation parameters of images at each time point relative to the reference volume are calculated, and samples with head movement exceeding 3mm or 3° are removed. The remaining samples are aligned to the reference volume through linear interpolation; Registration: Functional images and T1 structural images are rigidly registered to ensure spatial consistency; Segmentation: The tissue segmentation tool of SPM12 is used to separate gray matter, white matter, and cerebrospinal fluid for subsequent signal confusion regression; Normalization: The registered images are mapped to the MNI standard template and resampled to 3×3×3mm. 3 Voxel smoothing: Spatial smoothing is performed using a 6mm Gaussian kernel to enhance signal spatial correlation; Filtering: A 0.01-0.1Hz bandpass filter is used to retain neurophysiologically relevant signals and eliminate low-frequency drift and high-frequency noise.

[0024] It should be noted that by standardizing the entire process, the impact of interference factors such as head movement and differences in scanning layers on data quality was minimized, ensuring the consistency and comparability of data from different subjects and different scanning batches, thus laying a high-quality data foundation for subsequent feature extraction and model training. Furthermore, through a refined, end-to-end preprocessing strategy, a systematic improvement in data quality was achieved, solving the problems of residual noise and poor data consistency in traditional preprocessing methods.

[0025] Preferably, multi-template feature extraction is performed on the preprocessed data based on the first template and the second template to obtain the FC feature vector, including: The preprocessed data is divided into regions of interest using the first template and the second template, respectively, to obtain the first time series and the second time series. Calculate the Pearson correlation coefficients for the first and second time series respectively, and construct the FC matrix based on the Pearson correlation coefficients; Extract and exclude the lower triangular elements of the diagonal of the FC matrix to obtain the FC eigenvectors.

[0026] In the specific implementation process, step 200 loads the AAL template and the Power-264 template into the DPABI_V4.3 tool, and extracts the time series of all voxels within each ROI based on the region of interest (ROI) of the template, calculating the average value as the representative time series of that ROI. For each sample, the Pearson correlation coefficient between the time series of 116 ROIs of the AAL template and 264 ROIs of the Power-264 template is calculated to obtain a symmetric FC matrix. Specifically, if there are two variables (X) and (Y), their Pearson correlation coefficient (X) is calculated as follows: The formula for calculating ) is: ;in, The covariance represents the variance between variables (X) and (Y), used to describe the overall error between the two variables. and represents the standard deviation between variables (X) and (Y), used to measure the dispersion of the data for each variable, with values ​​limited to [-1, 1]. Then, eigenvectorization is performed, that is, extracting and excluding the lower triangular elements of the diagonal from the FC matrix to generate eigenvectors for the AAL template (6670 dimensions, 116 × 115 / 2) and the Power-264 template (34716 dimensions, 264 × 263 / 2).

[0027] It should be noted that by using a dual-template collaborative feature extraction method combining the AAL anatomical template and the Power-264 functional network template, both structural anatomical features and functional network topological features of the brain are simultaneously mined. The anatomical segmentation of the AAL template ensures accurate feature localization, while the functional network segmentation of the Power-264 template better reflects the synergistic characteristics of brain activity. This dual-template feature extraction method can comprehensively capture brain network abnormalities related to mTBI. Furthermore, the dual-template fusion extraction overcomes the limitations of single-template application and solves the problems of incomplete feature coverage and insufficient capture of brain network abnormalities in traditional methods.

[0028] Preferably, the FC feature vector is subjected to a three-layer progressive feature selection to obtain core discriminative features, including: Mutual information filtering is performed on the FC feature vectors to obtain the first-level filtered features; Random forest feature importance filtering is applied to the primary selection features to obtain secondary selection features; LASSO screening was performed on the secondary screening features to obtain the core discriminative features.

[0029] In the specific implementation process, step 300 uses the scikit-learn 1.2.2 toolkit in Python 3.8 to implement a three-layer feature selection process: mutual information (MI) selection, random forest (RF) selection, and LASSO selection. The specific process is as follows: Mutual Information (MI) Filtering: Define the MI calculation function, with the expression: ;in, As a feature, For classification purposes (mTBI=1, healthy controls=0). For joint probability distribution, , This represents a marginal probability distribution. After calculating the MI value between each feature and Y, features with MI > 0 are retained.

[0030] Random Forest (RF) Selection: Build RF model: n_estimators=100, other parameters are default; train the model using historical standard data and calculate feature importance, set a threshold of 0.0005, and retain features with importance higher than the threshold.

[0031] LASSO filtering: Define the LASSO objective function, the expression is: ;in, For the characteristic matrix, For the feature coefficient vector, For the classification target vector, For the sample size, For regularization parameters, (p is the number of features); traversed using 10-fold cross-validation. ∈[10-10,101] (1000 log interval values), and select the optimal value that minimizes the cross-validation error. At the same time, the coefficients Feature elimination with a value of 0.

[0032] It's important to note that the three-layer progressive feature selection process—mutual information (MI) for redundancy removal, random forest (RF) for importance assessment, and LASSO for strong sparsity—eliminates irrelevant and noisy features while retaining the core discriminative features of both templates. Through the synergistic effect of these three layers, MI ensures the relevance of features to the detection target, RF accurately assesses feature importance, and LASSO achieves sparsity compression of features. This avoids overfitting caused by high-dimensional data, accurately selects the most critical functional connectivity features, and balances feature effectiveness with model generalization ability. It avoids the problem of existing single-feature selection methods struggling to simultaneously balance feature relevance, importance, and sparsity, resulting in redundant features or missing key information. This achieves the complementary advantages of the three layers and solves the problems of low feature quality and the curse of dimensionality in traditional single-feature selection methods.

[0033] Preferably, the kernel function and hyperparameters of the classifier are optimized and validated based on the core discriminative features to obtain the trained classification model, including: The training set was obtained by performing 20 rounds of 5-fold cross-validation on the core discriminative features; The optimal parameters for the classifier's regularization and RBF kernel were optimized using grid search. The classifier is trained on the training set based on the optimal parameters to obtain the initial model; The initial model is validated across templates using core discriminative features to obtain a classification model.

[0034] In the specific implementation process, step 400 uses a radial basis function (RBF) kernel support vector machine (SVM) as the classifier. The SVM model is trained based on the core discriminative features selected from AAL and Power-264 templates, respectively. The robustness of the results is verified by calculating the overlapping key brain regions between the two templates. Specific steps include: Data partitioning: Repeat 5-fold cross-validation 20 times for the core discriminative features and take the average value to ensure the reliability of the training set.

[0035] Hyperparameter optimization: Grid search with 5-fold cross-validation is used to optimize the SVM regularization parameter C (10).-1 Up to 10 2 (10 logarithmic interval values) and RBF kernel parameters (10) -4 Up to 10 10 (50 log interval values), and the classification weights are set to "balanced" to balance the sample distribution, thus obtaining the optimal parameters.

[0036] Model Training: The SVM model is trained on the training set based on the optimal parameters. The RBF kernel function expression is as follows: ;in, , Given the feature vectors of two samples, The kernel parameter controls the range of influence of the samples. The expression for the SVM optimization problem is: ; ;in, For the weight vector, For bias terms, As slack variables, This is a regularization parameter used to balance model complexity and classification error.

[0037] Cross-template validation: Models were trained based on the core discriminative features selected from AAL and Power-264 templates, respectively, and the overlap of ROIs identified by the two models was calculated to verify the robustness of the results.

[0038] Model evaluation: Calculate the average of four metrics (ACC, SEN, SPE, and AUC) after 20 rounds of 5-fold cross-validation. Accuracy (ACC) is used to evaluate the classifier's ability to correctly predict classifications. The closer the accuracy is to 1, the better the classifier's classification performance. The formula for calculating accuracy is: Where TP represents the number of samples correctly predicted as positive by the model, TN represents the number of samples correctly predicted as negative by the model, FP represents the number of negative samples incorrectly predicted as positive by the model, and FN represents the number of positive samples incorrectly predicted as negative by the model. Sensitivity (SEN), also known as the true positive rate, is used to evaluate the classifier's ability to identify positive samples. It represents the proportion of samples that are actually positive out of all samples identified as positive by the model. Higher sensitivity indicates a lower probability of missed diagnoses. The formula for calculating sensitivity is: Specificity (SPE), also known as the true negative rate, is used to evaluate the classifier's ability to identify samples that are not diseased. It represents the proportion of truly disease-free samples among all samples identified as disease-free by the model. In other words, it's the probability that a sample that was originally negative will be detected as negative during the testing process. Higher specificity indicates a lower false positive rate. The formula for calculating specificity is: The area under the curve (AUC) is the area under the operating feature curve. The closer the AUC value is to 1, the better the classification performance of the classifier.

[0039] It should be noted that RBF and SVM can effectively capture the nonlinear relationships in high-dimensional features, while refined hyperparameter optimization further improves the model's classification accuracy and stability. Cross-template validation ensures that the results are not affected by a single template segmentation strategy, thereby significantly enhancing the model's robustness and clinical applicability. Furthermore, by optimizing the SVM model configuration and cross-template validation mechanism, the problems of poor model adaptability, insufficient robustness, and poor interpretability in traditional classification methods are addressed.

[0040] Preferably, the core discriminative features are functionally decomposed using a classification model to obtain functional connectivity pairs of a first template and a second template; the functional connectivity pairs include: a first brain region and a second brain region; Traverse all function connection pairs to obtain the node degree; Based on standard spatial coordinates, coordinate mapping is performed on the first and second templates to obtain the distribution of overlapping brain regions.

[0041] In the specific implementation process, the node degree in step 500 is the core indicator for measuring the importance of functional connectivity in brain regions. The calculation of node degree is based on the subset of functional connectivity features selected after three-layer combined feature selection. The specific process is as follows: taking the AAL template as an example, after screening through "MI redundancy removal → RF importance assessment → LASSO strong sparsity", the high-discriminative functional connectivity features are finally obtained, which constitute the core functional connectivity subset. For each pair of functional connections in this subset (denoted as ROI) i and ROI j The connection pairs were broken down to identify the two brain regions (ROIs) corresponding to each pair of connections. i and ROI j A counter is assigned to each Region of Interest (ROI). All functional connectivity pairs are iterated over, and the counter is incremented by 1 each time a connection involving that ROI occurs. The final counter value represents the node degree of that ROI, which is essentially the total number of other ROIs directly connected to it within the core functional connectivity feature subset. The node degree directly reflects the centrality of the brain region in the mTBI-related functional connectivity network. A higher node degree indicates more significant functional synergy abnormalities between that brain region and other key brain regions, and a greater contribution to mTBI detection.

[0042] The coordinate mapping process of the brain network is as follows: Based on the MNI standard spatial coordinates, the functional brain region coordinates of the Power-264 template are mapped to the AAL anatomical template. By determining the coordinate attribution, key ROIs common to both templates are selected. The specific steps are: extract all ROIs corresponding to the core functional connections under the Power-264 template, obtain the three-dimensional center coordinates (x, y, z) of each ROI, query the AAL template coordinates in the DPABI software according to the three-dimensional center coordinates of the Power-264-ROI, determine whether the center coordinates belong to the coordinate range of a certain AAL brain region, traverse all core ROIs of the Power-264 template, and summarize all brain regions that are determined to be homologous to the core ROIs of the AAL template through coordinate mapping, and finally obtain the cross-template overlapping ROIs. The overlapping ROIs were then further mapped to six core brain networks: the Default Mode Network (DMN), the Sensorimotor Network (SMN), the Subcortical Network (SCN), the Attention Network (AN), the Visual Network (VN), and the Cerebelum Network (CN).

[0043] It should be noted that by analyzing the distribution of overlapping regions of interest (ROIs) in each network, we can reveal the abnormal brain network function patterns related to mTBI, ensure the anatomical consistency of overlapping ROIs, effectively avoid the differences caused by different template segmentation strategies, and the selected overlapping ROIs are the common core regions of mTBI brain function abnormalities. Furthermore, it is clear that the brain function abnormalities in mTBI patients are not limited to a single network, but rather are the synergistic damage of multiple core networks.

[0044] Preferably, the first and second templates are mapped to coordinates based on standard spatial coordinates to obtain the distribution of overlapping brain regions, including: Obtain the three-dimensional center coordinates of the first brain region of the second template and the brain region coordinates of the second brain region of the second template; Determine whether the three-dimensional center coordinates of the first brain region of the second template and the three-dimensional center coordinates of the second brain region of the second template are within the range of the brain region coordinates of the first brain region and the second brain region of the first template, and summarize all the coordinates that pass the judgment to obtain the overlapping brain regions across templates. By mapping overlapping brain regions across templates to different core brain networks, the distribution of overlapping brain regions is obtained.

[0045] Preferably, the core brain network includes: default mode network, sensorimotor network, subcortical network, attention network, visual network, and cerebellar network.

[0046] The beneficial effects of this invention are as follows: 1) By integrating the AAL anatomical template and the Power-264 functional template, cross-template verification and key brain region overlap verification were achieved, which significantly improved the reliability and generalization ability of the detection results and solved the problems of result bias and applicability limitations of single template analysis. 2) By using a three-layer combination feature selection strategy of MI redundancy removal, RF importance screening, and LASSO strong sparsity, the shortcomings of traditional single feature screening methods are overcome. Through the three-step synergistic effect, noise and redundant information in high-dimensional rs-fMRI data are accurately removed, and core functional connectivity features with high discriminative power are efficiently screened out. 3) For the selected high-dimensional key features, the kernel function and hyperparameters of the SVM classifier are finely optimized, and the optimal parameter configuration is determined by combining grid search and cross-validation. This can accurately capture the nonlinear relationship between features and further improve the classification accuracy.

[0047] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0048] Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. Furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A brain information detection method based on multi-map and combined feature selection, characterized in that, Includes the following steps: The acquired rs-fMRI data were standardized to obtain preprocessed data; Multi-template feature extraction is performed on the preprocessed data based on the first template and the second template to obtain the FC feature vector; The FC feature vector is subjected to a three-layer progressive feature filtering process to obtain the core discriminative features; The kernel function and hyperparameters of the classifier are optimized and validated based on the core discriminative features to obtain the trained classification model. The core discriminative features are analyzed using the classification model to detect key ROIs and brain networks, resulting in the detection results, which include node degree and distribution of overlapping brain regions.

2. The brain information detection method based on multi-map and combined feature selection according to claim 1, characterized in that, The acquired rs-fMRI data were standardized to obtain preprocessed data, including: The rs-fMRI data were time-corrected to obtain time-corrected data; The time-corrected data is subjected to head movement correction and data alignment to obtain aligned data; The alignment data is rigidly registered and segmented with the T1 structural image to obtain registration data; The registration data is mapped to a standard template and then spatially smoothed and bandpass filtered to obtain the preprocessed data.

3. The brain information detection method based on multi-map and combined feature selection according to claim 1, characterized in that, Multi-template feature extraction is performed on the preprocessed data based on the first and second templates to obtain FC feature vectors, including: The preprocessed data is divided into regions of interest using the first template and the second template, respectively, to obtain a first time series and a second time series. Calculate the Pearson correlation coefficients for the first time series and the second time series respectively, and construct the FC matrix based on the Pearson correlation coefficients; Extract and exclude the lower triangular elements of the diagonal of the FC matrix to obtain the FC eigenvector.

4. The brain information detection method based on multi-map and combined feature selection according to claim 1, characterized in that, The FC feature vector is subjected to a three-layer progressive feature filtering process to obtain core discriminative features, including: Mutual information filtering is performed on the FC feature vectors to obtain first-level filtered features; Random forest feature importance filtering is applied to the first-level screening features to obtain the second-level screening features; The secondary screening features are subjected to LASSO screening to obtain the core discriminative features.

5. The brain information detection method based on multi-map and combined feature selection according to claim 1, characterized in that, Based on the core discriminative features, the kernel function and hyperparameters of the classifier are optimized and validated to obtain a trained classification model, including: The training set is obtained by performing 20 rounds of 5-fold cross-validation on the core discriminative features. The regularization parameters and RBF kernel parameters of the classifier are optimized by grid search to obtain the optimal parameters; Based on the optimal parameters, the classifier is trained on the training set to obtain an initial model; The initial model is validated across templates using the core discriminative features to obtain a classification model.

6. The brain information detection method based on multi-map and combined feature selection according to claim 5, characterized in that, The kernel function and hyperparameters of the classifier are optimized and validated based on the core discriminative features to obtain a trained classification model. The method also includes evaluating the classification model using evaluation metrics, which include accuracy, sensitivity, specificity, and area under the curve.

7. The brain information detection method based on multi-map and combined feature selection according to claim 1, characterized in that, The classification model is used to perform key ROI and brain network detection on the core discriminative features to obtain detection results, including: The core discriminative features are functionally decomposed using the classification model to obtain functional connectivity pairs of a first template and a second template; the functional connectivity pairs include: a first brain region and a second brain region. Traverse all the functional connection pairs to obtain the node degree; The first template and the second template are mapped using standard spatial coordinates to obtain the distribution of the overlapping brain regions.

8. The brain information detection method based on multi-map and combined feature selection according to claim 7, characterized in that, Based on standard spatial coordinates, coordinate mapping is performed on the first template and the second template to obtain the distribution of the overlapping brain regions, including: Obtain the three-dimensional center coordinates of the first brain region of the second template and the brain region coordinates of the second brain region of the second template; Determine whether the three-dimensional center coordinates of the first brain region of the second template and the three-dimensional center coordinates of the second brain region of the second template are within the range of brain region coordinates of the first brain region and the second brain region of the first template, and summarize all the coordinates that pass the judgment to obtain the overlapping brain regions across templates; The overlapping brain regions across templates are mapped to different core brain networks to obtain the distribution of the overlapping brain regions.

9. The brain information detection method based on multi-map and combined feature selection according to claim 8, characterized in that, The core brain networks include: default mode network, sensorimotor network, subcortical network, attention network, visual network, and cerebellar network.

Images