A three-dimensional visualization system and method for discriminating brain regions of AD based on machine learning and spatial statistics

Abstract

The application discloses a kind of AD discriminant brain area three-dimensional visualization system and method based on machine learning and space statistics, belong to the technical field of neurodegenerative disease diagnosis, the system includes: data preprocessing module, space statistical analysis and feature extraction module, machine learning modeling and optimization module and three-dimensional visualization rendering module;Data preprocessing module: for the original sMRI image is standardized, format conversion, normalization, and generate individual brain mask and intersection mask between groups;The present application realizes high-precision, robust multi-classification discrimination, the random forest and support vector machine model constructed by the present application, for the three-classification task of dementia, non-dementia, converter, the accuracy is high, the performance difference between training set and validation set is small, and excellent generalization ability is shown, which proves the effectiveness of the scheme in solving the problem of accurate staging.

Description

Technical Field

[0001] This invention belongs to the field of neurodegenerative disease diagnosis technology, specifically relating to a three-dimensional visualization system and method for AD discrimination brain regions based on machine learning and spatial statistics. Background Technology

[0002] Alzheimer's disease (AD) is a serious neurodegenerative disease, and its early and accurate identification and staging are of great significance for clinical intervention and prognosis improvement. Structural magnetic resonance imaging (sMRI), as a non-invasive imaging technique that can provide high-resolution brain anatomy, has become a key tool in assisting the diagnosis of AD.

[0003] Currently, sMRI-based methods for AD auxiliary diagnosis mainly include traditional methods that rely on manual analysis and automated methods based on machine learning. However, these methods all have certain limitations in practical applications. 1. Insufficient efficiency of traditional manual analysis methods: Traditional manual methods mainly rely on radiologists to visually evaluate sMRI images and manually delineate regions of interest (ROIs). The analysis results are limited by the doctor's subjective experience and have defects such as cumbersome analysis process, poor repeatability, and difficulty in standardization, which cannot meet the needs of large-scale clinical screening and accurate quantitative analysis. 2. Limited Feature Extraction and Classification Performance of Existing Automated Methods: Although machine learning-based methods have achieved automated analysis to some extent, their technical solutions have shortcomings. First, at the feature extraction level, most methods only use single morphological indicators (such as hippocampal volume) or local gray-scale statistical features, failing to fully and systematically integrate multi-dimensional information from brain regions, such as macroscopic morphological parameters (volume, surface area), microscopic gray-scale statistical distributions (skewness, kurtosis), and clinical prior information. This results in incomplete feature representation and an inability to effectively capture early and subtle pathological changes. Second, in classification tasks, existing methods are mostly focused on binary classification between AD and normal controls, lacking sufficient ability to distinguish the key category of "transformed mild cognitive impairment" which has higher clinical value, making it difficult to achieve accurate disease staging. 3. Lack of interpretability in the model decision-making process: Existing machine learning models are often regarded as "black boxes" because the anatomical or imaging basis of their classification decisions is not clearly revealed. Clinicians cannot trace and understand the key features and brain regions judged by the model, thereby reducing the credibility and clinical acceptance of the diagnostic results. 4. The analysis results are not presented intuitively: Current analysis results are mostly presented in the form of two-dimensional statistical charts or numerical lists, which lack a holistic and intuitive display of the anatomical location, spatial distribution and evolution of key discriminant brain regions in three-dimensional space. This is not conducive to clinicians quickly forming imaging insights and making individualized diagnoses.

[0004] Therefore, there is an urgent need for a technical solution that can overcome the above-mentioned defects, achieve multi-dimensional feature fusion of sMRI images, construct a high-precision AD multi-classification model, clarify the key basis of model decision-making, and provide intuitive three-dimensional visualization results, so as to improve the automation, accuracy and interpretability of AD assisted diagnosis. Summary of the Invention

[0005] To address the problems mentioned in the background, this invention provides a three-dimensional visualization system for AD discrimination brain regions based on machine learning and spatial statistics. This system fundamentally eliminates biases introduced by inconsistencies in the extent of individual brain regions, overcomes the limitations of single-type feature representation, comprehensively depicts AD brain structural abnormalities from microscopic to macroscopic perspectives, ensures robustness of the evaluation, and solves the problems of traditional two-dimensional presentation methods being unintuitive and lacking interpretability.

[0006] This invention also provides a method for three-dimensional visualization of brain regions for AD discrimination based on machine learning and spatial statistics.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a three-dimensional visualization system for AD discrimination brain regions based on machine learning and spatial statistics, the system comprising: a data preprocessing module, a spatial statistical analysis and feature extraction module, a machine learning modeling and optimization module, and a three-dimensional visualization rendering module; Data preprocessing module: used to standardize, convert, and normalize the raw sMRI images, and generate individual brain masks and intergroup intersection masks; Input: Raw T1-weighted MRI image (HDR / IMG format); Output: Standardized NIfTI format image (.nii.gz), with voxel resolution consistent with the original image; individual brain mask file (.nii format), smoothed with 6mm Gaussian; intergroup intersection mask file (.nii format), with a threshold of 0.8, retaining at least 80% of the voxels covered by the subjects. Spatial statistical analysis and feature extraction module: used to perform voxel-level intergroup statistical tests within a common mask space, and extract multi-dimensional feature sets from predefined ROIs and whole-brain regions; Input: Standardized sMRI images and intergroup intersection mask; Output: Statistical parametric plots (.nii, .gz, and png formats); a 2D array of 150 (samples) × 198 (features) feature matrix (rows correspond to samples, columns correspond to features), containing grayscale statistics, morphological, geometric properties, and clinical variables; Machine learning modeling and optimization module: used to standardize and oversample features, and to train and optimize random forest and support vector machine classification models; Input: A 150×198 feature matrix and its corresponding category labels; Output: Trained Random Forest (RF) and Support Vector Machine (SVM) model files (.pkl format); a list of feature importance rankings, associated with brain region labels or coordinates, for subsequent visualization mapping; 3D visualization rendering module: used to map statistically significant brain regions and key brain regions for model discrimination to standard brain space and generate interactive 3D maps; Input: Statistical parametric plot (.nii.gz), feature importance ranking and its corresponding brain region spatial coordinates, standard brain template; Output: Interactive 3D mind map (.html visualization file), supporting statistical significance or feature contribution of red-blue mapping, and can be rotated and viewed from multiple perspectives.

[0008] Furthermore, the data preprocessing module automatically generates individual brain masks based on the NiftiMasker tool in the Nilearn library; the spatial statistical analysis and feature extraction module locates differential brain regions based on the Welch's test and FDR correction at the voxel level; the machine learning modeling and optimization module builds models and optimizes hyperparameters based on the scikit-learn library; and the three-dimensional visualization rendering module uses the MNI152T1 brain template as a spatial reference and fsaverage5 cortical surface data as the rendering geometry.

[0009] Furthermore, the data preprocessing module, spatial statistical analysis and feature extraction module, and machine learning modeling and optimization module are developed based on the Python language and integrate the Nilearn, Scikit-learn, and Imbalanced-learn libraries; the 3D visualization rendering module integrates the Plotly.js library and supports interactive operation on the web.

[0010] Furthermore, a method for a three-dimensional visualization system of brain regions for AD discrimination based on machine learning and spatial statistics includes the following steps: Step 1: Data Preprocessing and Construction of Common Analysis Space: T1-weighted structural magnetic resonance imaging (sMRI) data and clinical information of the subjects were acquired. The original image format was converted to NIfTI format and Z-score intensity normalization was performed. Individual brain masks with 6mm Gaussian smoothing were generated for each subject. All individual brain masks were superimposed and an overlap threshold of 0.8 was set to generate inter-group intersection masks that retain at least 80% of the voxels covered by the subjects. The largest connected components were extracted to eliminate noise and a unified voxel analysis common space was constructed. Step 2: Spatial statistical analysis and differential brain region localization: Within the common space defined by the intergroup intersection mask, Welch's test was used to perform pairwise voxel-level comparisons between the dementia group, the non-dementia group, and the conversion group. The generated statistical parameter map was corrected for the false detection rate (FDR) (α=0.05) to identify statistically significant brain regions in the three group comparisons as candidate regions of interest (ROIs). Step 3: Multi-dimensional feature extraction: Based on prior knowledge of the brain regions involved in typical AD pathology and the significantly different brain regions obtained in Step 2, 20 key brain structures are selected as ROIs. Gray-scale statistical features, morphological and geometric features, group mask region features, and asymmetric features are extracted from each ROI. At the same time, global statistical and three-dimensional geometric features of the whole brain and differential brain regions are extracted. The subjects' age, gender, and clinical characteristics are integrated to construct a multi-dimensional feature set, and a feature matrix is ​​obtained through high variance screening. Step 4, Machine Learning Modeling and Optimization: Z-score standardization is performed on the feature matrix from Step 3 to eliminate dimensional differences; SMOTE oversampling technique is used during model training to address class imbalance caused by insufficient sample size in the transformation group; three-class classification models are constructed using random forest and support vector machine algorithms respectively; the hyperparameters of random forest are optimized through random search, and the hyperparameters of support vector machine are optimized through Bayesian optimization; hierarchical five-fold cross-validation is used to evaluate model performance and extract the importance of model features. Step 5: Feature Importance Analysis: The feature contribution of the random forest model is quantified by the Gini impurity decrease mean. The feature importance of the support vector machine model is ranked by analyzing the contribution of features to the predictive ability. The results of the two models are integrated to identify the key discriminative feature subset. Step Six: 3D Visualization Implementation: Map the brain regions with statistically significant differences from Step Two to the brain regions corresponding to the key features from Step Five onto the standard MNI152 brain template using 3D rendering technology, and generate an interactive 3D visualization atlas using 3D rendering technology.

[0011] Furthermore, the formula for calculating the Z-score intensity normalization in step one is as follows: in Original voxel values, and These are the mean and standard deviation of the image pixel values, respectively; the individual brain mask is automatically generated using the NiftiMasker tool in the Nilearn library with a 'whole-brain-template' strategy.

[0012] Furthermore, the predefined regions of interest (ROIs) in step two include 20 brain regions: the white matter of the left and right hemispheres, the cerebral cortex, the lateral ventricles, the thalamus, the caudate nucleus, the putamen, the globus pallidus, the hippocampus, the amygdala, and the nucleus accumbens.

[0013] Furthermore, the gray-scale statistical features in step three include nine indicators: mean, median, standard deviation, extreme values, 25th / 75th percentile, skewness, and kurtosis. The morphological and geometric features include seven indicators: volume, number of voxels, surface area, compactness, three-dimensional coordinates of the centroid, and the overlap ratio with the group of horizontal masks. The group mask region features include eight indicators: mean gray value of the group mask region, median gray value of the group mask region, standard deviation of gray value of the group mask region, minimum gray value of the group mask region, maximum gray value of the group mask region, volume of the group mask region, surface area of ​​the group mask region, and number of voxels of the group mask region. The asymmetry features include three indicators: asymmetry in volume between the left and right hemispheres, asymmetry in intensity between the left and right hemispheres, and asymmetry in surface area between the left and right hemispheres. The global features include eight indicators: mean gray level, standard deviation, skewness, kurtosis, sphericity, and equivalent diameter of the ellipsoid for the whole brain and differential brain regions.

[0014] Furthermore, the optimized hyperparameters for the random forest in step four include: max_depth=16, max_features=log2, min_samples_leaf=1, min_samples_split=2, n_estimators=476; The optimized hyperparameters of the support vector machine include: C=12.07, gamma=0.053, kernel=rbf; The performance evaluation metrics for the stratified five-fold cross-validation include accuracy, sensitivity for each category, specificity, and macro-average AUC value.

[0015] Furthermore, the 3D rendering in step six is ​​implemented using Python's nilearn library and Plotly.js library. The voxel statistics are mapped to the fsaverage5 cortical surface using the surface.vol_to_surf function. The statistics are encoded using RdBu color mapping. The initial statistical threshold is set to |t|>2.0. Hovering over or clicking on the brain region can display the precise MNI coordinates and statistical values / feature contribution.

[0016] Furthermore, in step six, the generation of the three-dimensional visualization map uses the surface.vol_to_surf function to map the statistical parameter map and key features to the fsaverage standard brain surface, extracts three-dimensional geometric data, creates an interactive three-dimensional mesh object using the Plotly.js library, and uses red and blue to map the statistical significance level or feature contribution, supporting multi-view rotation, scaling, hemisphere selection, and interactive querying of brain region coordinates and statistical values.

[0017] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention achieves high-precision and robust multi-class classification. The random forest and support vector machine model constructed in this invention has high accuracy for the three-class classification task of dementia, non-dementia, and conversion. The model has small performance differences between the training set and the validation set, demonstrating excellent generalization ability and proving the effectiveness of the solution in solving the problem of accurate staging.

[0018] 2. This invention provides interpretable decision-making basis. Feature importance analysis clearly reveals the key feature types that different models rely on. Random forest focuses on local gray-scale distribution, SVM focuses on overall morphology, and three-dimensional visualization results more clearly and intuitively show the spatial distribution of key discriminative brain regions and their evolution patterns at different stages of the disease, making the decision-making basis of the model transparent and understandable, effectively solving the model "black box" problem, and providing clinicians with direct imaging evidence.

[0019] 3. This invention provides a standardized and reproducible analysis process, from constructing a unified analysis space with inter-group intersection masks to a systematic multi-dimensional feature extraction specification, and then to a model training and evaluation process with embedded SMOTE, forming a complete set of operational standards. This helps to overcome the problems of arbitrary processes and difficult-to-reproduce results in traditional analysis, and improves the reliability and scalability of the method. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a flowchart of the data preprocessing process of the present invention; Figure 3 This is a statistical diagram showing the significant differences in brain regions between the dementia group and the non-dementia group in this invention. Figure 4 This is a statistical diagram showing the significant differences in brain regions between the dementia group and the conversion group in this invention. Figure 5 This is a statistical diagram showing the significant differences in brain regions between the non-dementia group and the conversion group in this invention. Figure 6 The ROC curves for various categories of the random forest model of this invention are shown below. Figure 7ROC curves for various types of support vector machine models in this invention. Figure 8 This is a graph showing the accuracy curves of the random forest and support vector machine at each fold of the present invention; Figure 9 This is a bar chart showing the feature importance of the random forest model of this invention; Figure 10 This is a bar chart showing the feature importance of the support vector machine model of the present invention. Figure 11 This is a brain region map showing the statistical differences between Demented and Nondemented regions as presented in this invention. Figure 12 This is a Demented vs. Converted statistical differential brain region map of the present invention; Figure 13 The Nondemented vs. Converted statistically differential brain region map of this invention Figure 14 This is a brain region distribution map showing the feature importance of the random forest model of this invention; Figure 15 This is a brain region distribution map showing the feature importance of the support vector machine model of this invention. Detailed Implementation

[0021] 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.

[0022] Example 1 This invention provides the following technical solution: a three-dimensional visualization system for AD discrimination brain regions based on machine learning and spatial statistics, the system comprising: a data preprocessing module, a spatial statistical analysis and feature extraction module, a machine learning modeling and optimization module, and a three-dimensional visualization rendering module; Data preprocessing module: used to standardize, convert, and normalize the raw sMRI images, and generate individual brain masks and intergroup intersection masks; Input: Raw T1-weighted MRI image (HDR / IMG format); Output: Standardized NIfTI format image (.nii.gz), with voxel resolution consistent with the original image; individual brain mask file (.nii format), smoothed with 6mm Gaussian; intergroup intersection mask file (.nii format), with a threshold of 0.8, retaining at least 80% of the voxels covered by the subjects. Spatial statistical analysis and feature extraction module: used to perform voxel-level intergroup statistical tests within a common mask space, and extract multi-dimensional feature sets from predefined ROIs and whole-brain regions; Input: Standardized sMRI images and intergroup intersection mask; Output: Statistical parametric plots (.nii, .gz, and png formats); a 2D array of 150 (samples) × 198 (features) feature matrix (rows correspond to samples, columns correspond to features), containing grayscale statistics, morphological, geometric properties, and clinical variables; Machine learning modeling and optimization module: used to standardize and oversample features, and to train and optimize random forest and support vector machine classification models; Input: A 150×198 feature matrix and its corresponding category labels; Output: Trained Random Forest (RF) and Support Vector Machine (SVM) model files (.pkl format); a list of feature importance rankings, associated with brain region labels or coordinates, for subsequent visualization mapping; 3D visualization rendering module: used to map statistically significant brain regions and key brain regions for model discrimination to standard brain space and generate interactive 3D maps; Input: Statistical parametric plot (.nii.gz), feature importance ranking and its corresponding brain region spatial coordinates, standard brain template; Output: Interactive 3D mind map (.html visualization file), supporting statistical significance or feature contribution of red-blue mapping, and can be rotated and viewed from multiple perspectives.

[0023] Furthermore, in this invention, the data preprocessing module automatically generates individual brain masks based on the NiftiMasker tool in the Nilearn library; the spatial statistical analysis and feature extraction module locates differential brain regions based on the Welch's test and FDR correction at the voxel level; the machine learning modeling and optimization module builds models and optimizes hyperparameters based on the scikit-learn library; and the three-dimensional visualization rendering module uses the MNI152T1 brain template as a spatial reference and fsaverage5 cortical surface data as the rendering geometry.

[0024] Furthermore, in this invention, the data preprocessing module, spatial statistical analysis and feature extraction module, and machine learning modeling and optimization module are developed based on the Python language and integrate the Nilearn, Scikit-learn, and Imbalanced-learn libraries; the 3D visualization rendering module integrates the Plotly.js library and supports interactive operation on the Web.

[0025] Example 2 Based on the same inventive concept, this embodiment provides a method for a three-dimensional visualization system of brain regions for AD discrimination based on machine learning and spatial statistics. The principle of solving the problem is similar to that of the aforementioned three-dimensional visualization system for AD discrimination based on machine learning and spatial statistics, and the repetitions will not be repeated.

[0026] This embodiment provides a method for a three-dimensional visualization system of brain regions for AD discrimination based on machine learning and spatial statistics, including the following steps: Step 1: Data Preprocessing and Construction of Common Analysis Space: T1-weighted structural magnetic resonance imaging (sMRI) data and clinical information of the subjects were acquired. The original image format was converted to NIfTI format and Z-score intensity normalization was performed. Individual brain masks with 6mm Gaussian smoothing were generated for each subject. All individual brain masks were superimposed and an overlap threshold of 0.8 was set to generate inter-group intersection masks that retain at least 80% of the voxels covered by the subjects. The largest connected components were extracted to eliminate noise and a unified voxel analysis common space was constructed. Step 2: Spatial statistical analysis and differential brain region localization: Within the common space defined by the intergroup intersection mask, Welch's test was used to perform pairwise voxel-level comparisons between the dementia group, the non-dementia group, and the conversion group. The generated statistical parameter map was corrected for the false detection rate (FDR) (α=0.05) to identify statistically significant brain regions in the three group comparisons as candidate regions of interest (ROIs). Step 3: Multi-dimensional feature extraction: Based on prior knowledge of the brain regions involved in typical AD pathology and the significantly different brain regions obtained in Step 2, 20 key brain structures are selected as ROIs. Gray-scale statistical features, morphological and geometric features, group mask region features, and asymmetric features are extracted from each ROI. At the same time, global statistical and three-dimensional geometric features of the whole brain and differential brain regions are extracted. The subjects' age, gender, and clinical characteristics are integrated to construct a multi-dimensional feature set, and a feature matrix is ​​obtained through high variance screening. Step 4, Machine Learning Modeling and Optimization: Z-score standardization is performed on the feature matrix from Step 3 to eliminate dimensional differences; SMOTE oversampling technique is used during model training to address class imbalance caused by insufficient sample size in the transformation group; three-class classification models are constructed using random forest and support vector machine algorithms respectively; the hyperparameters of random forest are optimized through random search, and the hyperparameters of support vector machine are optimized through Bayesian optimization; hierarchical five-fold cross-validation is used to evaluate model performance and extract the importance of model features. Step 5: Feature Importance Analysis: The feature contribution of the random forest model is quantified by the Gini impurity decrease mean. The feature importance of the support vector machine model is ranked by analyzing the contribution of features to the predictive ability. The results of the two models are integrated to identify the key discriminative feature subset. Step Six: 3D Visualization Implementation: Map the brain regions with statistically significant differences from Step Two to the brain regions corresponding to the key features from Step Five onto the standard MNI152 brain template using 3D rendering technology, and generate an interactive 3D visualization atlas using 3D rendering technology.

[0027] Furthermore, in this invention, the formula for calculating the Z-score intensity normalization in step one is as follows: in Original voxel values, and These are the mean and standard deviation of the image pixel values, respectively; individual brain masks are automatically generated using the NiftiMasker tool in the Nilearn library with a 'whole-brain-template' strategy.

[0028] Furthermore, in this invention, the predefined regions of interest (ROIs) in step two include 20 brain regions: the white matter of the left and right hemispheres, the cerebral cortex, the lateral ventricles, the thalamus, the caudate nucleus, the putamen, the globus pallidus, the hippocampus, the amygdala, and the nucleus accumbens.

[0029] Furthermore, in this invention, the gray-scale statistical features in step three include nine indicators: mean, median, standard deviation, extreme values, 25th / 75th percentile, skewness, and kurtosis. Morphological and geometric features include seven indicators: volume, number of voxels, surface area, compactness, three-dimensional coordinates of the centroid, and the overlap ratio with the group of horizontal masks. The characteristics of the group mask region include eight indicators: mean gray value of the group mask region, median gray value of the group mask region, standard deviation of gray value of the group mask region, minimum gray value of the group mask region, maximum gray value of the group mask region, volume of the group mask region, surface area of ​​the group mask region, and number of voxels of the group mask region. The asymmetry features include three indicators: asymmetry in volume between the left and right hemispheres, asymmetry in intensity between the left and right hemispheres, and asymmetry in surface area between the left and right hemispheres. The global features include eight indicators: mean gray level, standard deviation, skewness, kurtosis, sphericity, and equivalent diameter of the ellipsoid in the whole brain and differential brain regions.

[0030] Furthermore, in this invention, the optimized hyperparameters for the random forest in step four include: max_depth=16, max_features=log2, min_samples_leaf=1, min_samples_split=2, n_estimators=476; The optimized hyperparameters for the support vector machine include: C=12.07, gamma=0.053, kernel=rbf; The performance evaluation metrics for stratified five-fold cross-validation include accuracy, sensitivity for each category, specificity, and macro-average AUC.

[0031] Furthermore, in this invention, the three-dimensional rendering in step six is ​​implemented using Python's nilearn library and Plotly.js library. The voxel statistics are mapped to the fsaverage5 cortical surface using the surface.vol_to_surf function, and the statistics are encoded using RdBu color mapping. The initial statistical threshold is set to |t|>2.0. Hovering over or clicking on the brain region can display the precise MNI coordinates and statistical values / feature contribution.

[0032] Furthermore, in step six of this invention, the generation of the three-dimensional visualization map uses the surface.vol_to_surf function to map the statistical parameter map and key features to the fsaverage standard brain surface, extracts three-dimensional geometric data, creates an interactive three-dimensional mesh object using the Plotly.js library, and uses red and blue to map the statistical significance level or feature contribution, supporting multi-view rotation, scaling, hemisphere selection, and interactive querying of brain region coordinates and statistical values.

[0033] Example 3 Furthermore, this embodiment of the invention provides a specific method for a three-dimensional visualization system of brain regions for AD discrimination based on machine learning and spatial statistics; the method includes the following steps: Step 1: Data Preprocessing and Construction of Common Analysis Space S1 data sources and subject information sources: The data in this application comes from the publicly available OASIS-2 longitudinal dataset, which mainly studies the longitudinal changes in brain structure over time in non-demented and demented elderly people. A total of 150 subjects aged between 60 and 96 years were included. All subjects underwent multiple visits, with each visit at least one year apart, for a total of 373 imaging scan sessions. Each scan session included 3 to 4 separate T1-weighted magnetic resonance imaging scans, and the corresponding clinical and demographic information was recorded. In order to control for possible variations due to multiple scans and to ensure that the data of all subjects were at comparable time points, this application uniformly used the scan data from the first visit and baseline clinical information, including variables such as age, gender, and group information. The subjects were divided into three groups: (1) Dementia group: clinically diagnosed Alzheimer's disease dementia patients; (2) Non-dementia group: healthy controls with normal cognitive function; (3) Conversion group: subjects who were converted from an initial non-dementia state to mild cognitive impairment or dementia during the follow-up period. All images underwent strict quality control to ensure the reliability of subsequent analysis. S2 data preprocessing and intergroup intersection mask generation: This application employs a standardized data preprocessing procedure (such as...). Figure 2 (as shown); First, the HDR / IMG format of the raw data was converted to the NIfTI format, which is common for neuroimaging analysis, and the T1-weighted image data of each subject was Z-score normalized (Equation (1)) to minimize the impact of inter-scanner differences and inter-individual global intensity differences on subsequent analysis. (1) in Original voxel values, and These are the mean and standard deviation of the pixel values ​​in the image, respectively. Based on this, the NiftiMasker tool in the Nilearn library was used to automatically generate an individual brain mask for each subject using a 'whole-brain-template' strategy. At the same time, a 6mm full-width half-height Gaussian smoothing kernel was used to further improve the continuity and stability of the mask. To ensure that subsequent voxel-level statistical analysis and feature extraction can be performed within a unified spatial range, this application further superimposed the brain masks of all individuals to generate an inter-group intersection mask. The threshold was set to 0.8, that is, to retain at least 80% of the voxels that are commonly covered by the subjects' brains, and to screen out the largest connected components to eliminate the influence of discrete noise. Finally, the quality of the mask generation was verified by visualization: the binarized individual brain mask was superimposed on the original T1 image in a semi-transparent form. With the help of orthogonal view and high-contrast color mapping, the degree of fit between the brain region defined by the mask and the original anatomical structure was intuitively evaluated, confirming its extraction accuracy and providing a reliable basis for subsequent localization analysis. Step 2: Spatial statistical analysis and differential brain region localization: To identify patterns of brain structural changes associated with the progression of Alzheimer's disease, this application conducted intergroup comparisons among three groups of subjects: a dementia group, a non-dementia group, and a conversion group. The comparison between the dementia and non-dementia groups was used to identify typical AD atrophy patterns; the comparison between the conversion and non-dementia groups was used to identify early brain structural changes that could predict the progression to dementia; and the comparison between the dementia and conversion groups was used to reveal brain structural evolution during disease progression. Within the common area defined by the intergroup intersection mask, a two-sample t-test (Welch's-test) at the voxel level was used for statistical analysis, and the false discovery rate of the generated statistical parameter plots was calculated. (FDR) correction defines the brain regions that are significant after FDR correction in the three comparisons as candidate regions of interest (ROIs). These brain regions include both the typical atrophic areas of AD and the specific changes in the disease transformation process, providing rich discriminative features for the three-classification task. Finally, orthogonal views are used to simultaneously display the statistical parameter graphs of the three comparisons, intuitively presenting the spatial distribution characteristics and overlap of brain structural differences at different clinical stages, ensuring that the extracted features can effectively distinguish different clinical stages of Alzheimer's disease, and providing biologically significant discriminative brain regions for subsequent three-classification machine learning models. Step 3: Multi-dimensional feature extraction: To construct a feature set that can effectively distinguish different clinical stages of Alzheimer's disease (AD), this application constructs a multi-dimensional feature set based on preprocessed structural magnetic resonance images, which integrates local brain region characteristics, global statistical attributes and clinical information, aiming to quantify changes in brain structure and function from multiple perspectives such as radiomics and clinical phenotypes. Based on prior knowledge of the typical brain regions involved in Alzheimer's disease, this application selects 20 key brain structures, including the bilateral hippocampus, amygdala, thalamus, basal ganglia nuclei (caudate nucleus, putamen, globus pallidus), nucleus accumbens, as well as cerebral white matter, cortex, and lateral ventricles, as regions of interest (ROIs). Two types of quantitative features are systematically extracted from each ROI: one type is gray-scale statistical features, including nine indicators such as mean, median, standard deviation, extreme values, quantiles (25th and 75th percentiles), skewness, and kurtosis, to describe the central tendency, dispersion, and morphological characteristics of gray matter intensity distribution within the ROI, thereby reflecting potential microscopic changes in tissue density or composition; the other type is morphological and geometric features, including seven indicators such as volume, number of voxels, surface area, compactness, three-dimensional coordinates of the centroid, and the overlap ratio with the group-level mask, to quantify the structural attributes of brain regions at a macroscopic scale, such as the degree of atrophy, shape regularity, and spatial variation. To supplement the lack of local ROI information, this application further extracts statistical features of the whole brain and differential regions, including five global indicators: mean gray level of the whole brain, mean gray level of differential regions, standard deviation, skewness, and kurtosis. This provides the model with overall background information that is synergistic with local characteristics. At the same time, for the significantly differential brain regions identified in the statistical comparison, three-dimensional geometric properties such as sphericity, compactness, and ellipsoidal equivalent diameter are additionally calculated as shape features to enhance the ability to describe disease-specific morphological changes. Clinical variables, including age and gender, are also included to integrate demographic and clinical diagnostic information that is closely related to cognitive state and disease progression, thereby enhancing the applicability and interpretability of the model in real medical scenarios. Step 4: Machine Learning Modeling and Optimization To eliminate the impact of differences in the units of features on model training, all numerical features are Z-score standardized to have a mean of 0 and a standard deviation of 1. Meanwhile, to address the class imbalance problem in classification tasks (the number of samples in the Converted group is small), SMOTE is used to oversample the training set data during model training to improve the model's ability to identify a minority of classes. This application uses two algorithms, random forest and support vector machine, to construct a three-class classification model. To ensure optimal model performance, hyperparameter optimization was performed on both algorithms: For the random forest model, a random search strategy is used to optimize key parameters such as the number of decision trees, maximum depth, minimum number of split samples, minimum number of leaf node samples, and maximum number of features. For the support vector machine model, the Bayesian optimization method is used to search for the optimal regularization parameter C, kernel function type, and kernel function coefficient gamma. The model evaluation adopts a hierarchical five-fold cross-validation strategy to ensure that the proportion of classes in each subset is consistent with the overall dataset. The final model performance is comprehensively evaluated by accuracy, sensitivity and specificity of each class, and macro average AUC value. Step 5: Feature Importance Analysis and 3D Visualization Implementation To improve the interpretability of classification models and identify key discriminative features, this application comprehensively evaluates the feature importance of random forest and support vector machine models. Random forest uses the average value based on the decrease of Gini impurity to quantify the feature contribution, while support vector machine ranks the features by analyzing their contribution to the model's predictive ability and generates a feature importance map. By integrating the importance evaluation results of the two models, a stable subset of features that is most discriminative in distinguishing between dementia, non-dementia, and conversion patients is identified. Based on this, the brain regions with significant differences and the image features with the highest importance of features found in the statistical comparison are mapped to the standard brain space, and their spatial distribution is visualized by three-dimensional rendering technology. This provides a panoramic view of the anatomical location and spatial relationship of key brain regions, overcomes the limitations of traditional two-dimensional slice display, and enhances the interpretability of the results. S1 Study Subjects and Data Overview: This application ultimately included 150 participants from the OASIS-2 dataset. Their initial visit data, after rigorous quality control and preprocessing, were divided into three groups: Demented (64 participants), Nondemented (72 participants), and Converted (14 participants). Basic information of all participants and inter-group comparisons are shown in Table 1. Table 1. Basic information of all participants and intergroup comparison results: Note: Age data were analyzed using one-way ANOVA; gender composition was analyzed using a chi-square test.

[0034] As shown in Table 1, one-way ANOVA showed that there was no significant difference in age distribution among the three groups of subjects (F=0.385, p>0.05), while the chi-square test showed a significant difference in gender distribution among the three groups of subjects (χ²=10.26, p<0.05). The cohort in this application is well matched in age and has good representativeness. S2 intergroup differential brain region analysis: To locate significantly different brain regions in different clinical stages of Alzheimer's disease, this application conducted voxel-level intergroup comparisons of the dementia group, conversion group, and non-dementia group based on intergroup intersection masking. Welch's test was used, combined with false discovery rate correction (α=0.05), to identify significantly different brain regions in the three groups. Their spatial distribution is clearly presented through orthogonal views. S2.1 Dementia Group vs. Non-Dementia Group: Confirmation of Typical Atrophy Pattern in Alzheimer's Disease (AD): like Figure 3 As shown, the orthogonal view reveals a large area of ​​light blue, indicating extensive gray matter atrophy in dementia patients. These areas are mainly distributed in the core nodes of the default mode network (DMN), especially the precuneus and posterior cingulate cortex (most obvious in the sagittal plane), as well as the frontal pole (visible in the coronal and transverse planes). At the same time, the lingual gyrus region (visible in the transverse plane) also shows significant atrophy. This atrophy pattern, mainly in the midline structures and posterior cortex, visually demonstrates the typical neuropathological changes of AD in the images. S2.2 Dementia Group vs. Transformation Group: Cortical Expansion During Disease Progression Figure 4The orthogonal views show that, compared with the conversion group, the brain atrophy in the dementia group has further expanded. In addition to the persistent precuneus atrophy, the precentral gyrus and postcentral gyrus (distributed along the central sulcus in the sagittal and transverse planes) have become new and significant atrophic areas. Extensive yellow-red signals in the sensorimotor cortex are visible in the images, indicating that during the progression of the disease, the atrophy has expanded from the higher association cortex to the primary motor sensorimotor cortex. All three planes show persistent atrophic signals in the prefrontal cortex, reflecting the progressive decline in executive function. S2.3 Non-dementia group vs. conversion group: Key brain regions predicting early conversion: Figure 5 The orthogonal view revealed the most revealing findings. The images showed that in the preclinical stage, the frontal pole had already atrophied (especially in the coronal and sagittal planes), while the precuneus and upper lateral occipital cortex also showed early changes. Notably, the precentral gyrus and postcentral gyrus had already begun to show atrophy signals at this stage. This finding was clearly visible on the transverse images. The extensive atrophy pattern, mainly in the frontoparietal system, appeared as a continuous yellow-red area on the images, providing important radiological evidence for the early identification of AD. The above results clearly reveal the spatiotemporal evolution trajectory of brain structural damage in Alzheimer's disease (AD), from... Figures 2 to 4 It can be observed that the atrophy pattern progresses from extensive involvement of the prefrontal-parietal lobes in the early stage of transformation to the typical AD pattern characterized by sensorimotor cortex invasion in the dementia stage. The significant brain regions corrected by FDR not only provide reliable imaging evidence for the staging of AD, but also lay a solid foundation for the subsequent construction of a high-precision classification model. S3 classification model performance: Based on 198 high-variance features selected after fusion, Random Forest (RF) and Support Vector Machine (SVM) models were constructed to perform three-class classification tasks (Demented, Nondemented, Converted). To alleviate the class imbalance problem in the training data, SMOTE oversampling technology was introduced during the model training stage. The hyperparameters of the Random Forest were optimized through a random search strategy, and the optimal parameter configuration of the Support Vector Machine was determined using a Bayesian optimization method. The specific optimization results are shown in Table 2. Table 2. Model hyperparameter optimization results: ROC curve of cross-validation ( Figure 6 Here is the ROC curve for the random forest. Figure 7 (ROC curve of support vector machine) and accuracy plots for each fold ( Figure 8 The following tables are used to comprehensively evaluate model performance: Table 3. Model classification performance evaluation results (hierarchical five-fold cross-validation): The random forest model achieved an average accuracy of 97.25% (±1.92%) on the validation set, while the support vector machine model achieved an average accuracy of 96.91% (±2.35%). Both models demonstrated excellent discriminative ability across all groups. Specifically, the random forest model achieved AUC values ​​of 0.998, 0.996, and 0.975 for the Demented, Nondemented, and Converted classes, respectively; the support vector machine model achieved AUC values ​​higher than 0.995 for all classes. The overfitting analysis results show that the performance gap between random forest and support vector machine on the training and validation sets is small, with relative overfitting rates of 2.75% and 3.09%, respectively. This indicates that neither model shows significant overfitting and both have good generalization ability. S4 Feature Importance Analysis: Feature importance analysis was performed on random forest and support vector machine models respectively to reveal the key basis of model decision-making and enhance interpretability. Random forest uses the average value based on the decrease of Gini impurity to evaluate the feature contribution, while support vector machine ranks the features by analyzing their contribution to the model's predictive ability. Figure 9 The study presented the top 20 most important features in the random forest model. The results showed that the features that contributed the most to classification included: right putamen kurtosis, 25th percentile of right white matter, 25th percentile of left white matter, right globus pallidus kurtosis, and right putamen skewness. The results indicated that the distribution morphology of the basal ganglia region (especially the right putamen and globus pallidus) and the gray-scale distribution characteristics of the bilateral white matter are of outstanding importance in AD discrimination. In addition, hippocampal-related features (such as left hippocampal kurtosis and standard deviation) are also among the important features, but their contribution is relatively lower than that of the aforementioned basal ganglia and white matter features. The feature importance ranking results of the support vector machine model are as follows: Figure 10 As shown, its key features differ significantly from those of the random forest model. The most important features are the three-dimensional morphological parameters, including the volume, surface area, compactness, and equivalent diameter of the ellipsoid in the three spatial dimensions of the differential brain regions. The importance values ​​of these morphological features are higher than those of the gray-level statistical features, indicating that the discrimination decision of the SVM model is highly dependent on the macroscopic three-dimensional geometric properties of the differential brain regions. This result highlights the sensitivity of the SVM algorithm to the overall morphology and spatial configuration of the brain structure, which is significantly different from the random forest model that focuses on the local gray-level distribution characteristics. The combined feature importance analysis results of the two models reveal that the gray-scale distribution characteristics of bilateral brain white matter are jointly identified as key discriminant factors. Furthermore, the two models emphasize different types of features: Random Forest focuses more on the gray-scale statistical distribution characteristics of local brain regions (such as kurtosis, skewness, and quantiles), while Support Vector Machine relies more on global morphological parameters (such as volume, surface area, and spatial dimension). The complementarity of the two models in feature selection indicates that local microstructure and overall macromorphology have a synergistic effect in AD discrimination, thus verifying the necessity and effectiveness of the multi-feature fusion strategy. That is, by integrating multi-dimensional information, AD-related brain structural changes can be more comprehensively characterized, providing a key basis for the high performance of the model. S5-based 3D visualization of brain regions: To visually represent the spatial distribution characteristics of key brain regions for Alzheimer's disease (AD), this application uses voxel-level statistical comparison results and feature importance analysis to perform three-dimensional visualization of brain regions with statistical differences and brain regions with feature importance, respectively. Figure 11 , 12 Figures 1 and 13 present interactive 3D visualizations of brain regions with statistical differences in three sets of comparisons (Demented vs. Nondemented, Demented vs. Converted, and Nondemented vs. Converted). The images are rendered on a standard MNI152 brain template and the statistical significance level is mapped using red and blue to clearly show the specific anatomical location and spatial distribution pattern of the differences in brain structure between the groups. Compared with traditional two-dimensional slices, it presents the spatial configuration of brain regions in a panoramic way, overcomes the limitations of two-dimensional projection, and reveals the evolutionary trajectory of AD-related brain atrophy more intuitively. Figure 14 and 15 This study presents a 3D brain map based on feature importance analysis. The key features of the random forest model are mainly located in the basal ganglia (such as the right putamen and globus pallidus) and white matter regions of the brain, reflecting the importance of local gray-scale statistical characteristics in AD discrimination. In contrast, the important features of the support vector machine model are widely distributed in brain regions with significant morphological changes, highlighting the dominant role of overall geometry in SVM decision-making. The 3D feature importance brain map reveals the complementarity of different models in feature selection from a spatial perspective, providing interpretable visual evidence for understanding the discrimination mechanism of machine learning models.

[0035] Example 4 This embodiment provides a clinical application of a three-dimensional visualization system and method for AD discrimination of brain regions based on machine learning and spatial statistics: Hospital Neurology / Radiology Department Auxiliary Diagnostic System: Application method: As an intelligent analysis plugin or independent diagnostic workstation of the medical institution's PACS system, it can be integrated into the existing medical imaging workflow; Specific applications: After obtaining the patient's sMRI data, doctors can use this system for automated analysis to obtain comprehensive diagnostic references, including disease classification results, three-dimensional visualization atlases of key brain regions, and feature importance reports. This assists in the early screening, differential diagnosis, and staging of Alzheimer's disease (AD), compensates for the subjective differences among doctors in recognizing subtle changes in brain structure, and improves diagnostic efficiency and accuracy. It is particularly helpful in discovering early and atypical AD cases.

[0036] Brain health screening platform for physical examination centers and health management institutions: Application method: As a service in high-end physical examination or brain health assessment packages, it is aimed at people with a family history of AD or those at risk of subjective cognitive decline; Specific applications: Rapidly analyze the sMRI data of examinees to assess the probability of them belonging to the categories of "normal", "high risk (converting)" or "abnormal (AD)", and generate easy-to-understand brain health reports and 3D visualization results for risk assessment and health management guidance. This enables non-invasive, rapid and low-cost preliminary screening of large populations and provides a basis for early intervention for high-risk groups.

[0037] Example 5 This embodiment provides an application of a three-dimensional visualization system and method for AD discrimination of brain regions based on machine learning and spatial statistics in scientific research and drug development: Neuroimaging analysis tools used by research institutions: Application: As a software tool for biomarker discovery and disease mechanism research, it is used by researchers in fields such as neuroscience and medical imaging. Specific applications: used to analyze image data from different cohorts, verify new scientific hypotheses, accurately quantify the trajectory of brain structural evolution during AD progression, and identify new imaging biomarkers, thereby improving the efficiency and reliability of scientific research.

[0038] Clinical trial evaluation and patient stratification tools for pharmaceutical companies: Application method: Integrate into the data management system of new drug clinical trials as an auxiliary tool for efficacy evaluation and patient enrollment; Specific applications: Patient screening: Accurately identify early-stage Alzheimer's disease (AD) patients in the "conversion group" who meet the inclusion criteria; Efficacy assessment: By comparing changes in differential brain regions before and after treatment and changes in model scores, objective and quantitative imaging evidence is provided for drug efficacy, improving the homogeneity of patients enrolled in clinical trials, and identifying imaging biomarkers that are more sensitive to treatment response, thus accelerating the new drug development process.

[0039] Example 6 This embodiment provides an application of a three-dimensional visualization system and method for AD discrimination of brain regions based on machine learning and spatial statistics in education and social services: Interactive teaching systems for medical education and training: Application: Developed into teaching software or simulation systems for medical schools and residency training; Specific applications: Through interactive operations, students can intuitively learn the three-dimensional spatial distribution of typical and atypical atrophic brain regions in Alzheimer's disease (AD), understand the discrimination logic of machine learning models, improve their ability to interpret images, and transform abstract medical knowledge into intuitive and vivid three-dimensional models, thus overcoming teaching difficulties.

[0040] Cloud service platform for telemedicine and community health service centers: Application method: Provide technical support to hospitals or community health service centers in areas with scarce medical resources in the form of cloud services (SaaS); Specific applications: Primary hospitals can upload sMRI data to the cloud and obtain professional analysis reports generated by this system, realizing the downward flow of high-quality diagnostic resources, narrowing the gap in AD diagnosis levels between different regions, and improving the equalization of public health services.

[0041] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A three-dimensional visualization system for AD discrimination brain regions based on machine learning and spatial statistics, characterized in that: The system includes: a data preprocessing module, a spatial statistical analysis and feature extraction module, a machine learning modeling and optimization module, and a 3D visualization rendering module; Data preprocessing module: used to standardize, convert, and normalize the raw sMRI images, and generate individual brain masks and intergroup intersection masks; Input: Raw T1-weighted MRI image (HDR / IMG format); Output: Standardized NIfTI format image (.nii.gz), with voxel resolution consistent with the original image; individual brain mask file (.nii format), smoothed with 6mm Gaussian; intergroup intersection mask file (.nii format), with a threshold of 0.8, retaining at least 80% of the voxels covered by the subjects. Spatial statistical analysis and feature extraction module: used to perform voxel-level intergroup statistical tests within a common mask space, and extract multi-dimensional feature sets from predefined ROIs and whole-brain regions; Input: Standardized sMRI images and intergroup intersection mask; Output: Statistical parametric plots (.nii, .gz, and png formats); a 2D array of 150 (samples) × 198 (features) feature matrix (rows correspond to samples, columns correspond to features), containing grayscale statistics, morphological, geometric properties, and clinical variables; Machine learning modeling and optimization module: used to standardize and oversample features, and to train and optimize random forest and support vector machine classification models; Input: A 150×198 feature matrix and its corresponding category labels; Output: Trained Random Forest (RF) and Support Vector Machine (SVM) model files (.pkl format); a list of feature importance rankings, associated with brain region labels or coordinates, for subsequent visualization mapping; 3D visualization rendering module: used to map statistically significant brain regions and key brain regions for model discrimination to standard brain space and generate interactive 3D maps; Input: Statistical parametric plot (.nii.gz), feature importance ranking and its corresponding brain region spatial coordinates, standard brain template; Output: Interactive 3D mind map (.html visualization file), supporting statistical significance or feature contribution of red-blue mapping, and can be rotated and viewed from multiple perspectives.

2. The three-dimensional visualization system for AD discrimination brain regions based on machine learning and spatial statistics as described in claim 1, characterized in that: The data preprocessing module uses the NiftiMasker tool from the Nilearn library to automatically generate individual brain masks; the spatial statistical analysis and feature extraction module uses the Welch's test and FDR correction at the voxel level to locate differential brain regions; the machine learning modeling and optimization module uses the scikit-learn library to build models and optimize hyperparameters; and the 3D visualization rendering module uses the MNI152T1 brain template as a spatial reference and fsaverage5 cortical surface data as the rendering geometry.

3. The three-dimensional visualization system for AD discrimination brain regions based on machine learning and spatial statistics as described in claim 1, characterized in that: The data preprocessing module, spatial statistical analysis and feature extraction module, and machine learning modeling and optimization module are developed based on the Python language and integrate the Nilearn, Scikit-learn, and Imbalanced-learn libraries; the 3D visualization rendering module integrates the Plotly.js library and supports interactive operation on the web.

4. A method for a three-dimensional visualization system of brain regions for AD discrimination based on machine learning and spatial statistics, as described in any one of claims 1-3, characterized in that: Includes the following steps: Step 1: Data Preprocessing and Construction of Common Analysis Space: T1-weighted structural magnetic resonance imaging (sMRI) data and clinical information of the subjects were acquired. The original image format was converted to NIfTI format and Z-score intensity normalization was performed. Individual brain masks with 6mm Gaussian smoothing were generated for each subject. All individual brain masks were superimposed and an overlap threshold of 0.8 was set to generate inter-group intersection masks that retain at least 80% of the voxels covered by the subjects. The largest connected components were extracted to eliminate noise and a unified voxel analysis common space was constructed. Step 2: Spatial statistical analysis and differential brain region localization: Within the common space defined by the intergroup intersection mask, Welch's test was used to perform pairwise voxel-level comparisons between the dementia group, the non-dementia group, and the conversion group. The generated statistical parameter map was corrected for the false detection rate (FDR) (α=0.05) to identify statistically significant brain regions in the three group comparisons as candidate regions of interest (ROIs). Step 3: Multi-dimensional feature extraction: Based on prior knowledge of the brain regions involved in typical AD pathology and the significantly different brain regions obtained in Step 2, 20 key brain structures are selected as ROIs. Gray-scale statistical features, morphological and geometric features, group mask region features, and asymmetric features are extracted from each ROI. At the same time, global statistical and three-dimensional geometric features of the whole brain and differential brain regions are extracted. The subjects' age, gender, and clinical characteristics are integrated to construct a multi-dimensional feature set, and a feature matrix is ​​obtained through high variance screening. Step 4, Machine Learning Modeling and Optimization: Z-score standardization is performed on the feature matrix from Step 3 to eliminate dimensional differences; SMOTE oversampling technique is used during model training to address class imbalance caused by insufficient sample size in the transformation group; three-class classification models are constructed using random forest and support vector machine algorithms respectively; the hyperparameters of random forest are optimized through random search, and the hyperparameters of support vector machine are optimized through Bayesian optimization; hierarchical five-fold cross-validation is used to evaluate model performance and extract the importance of model features. Step 5: Feature Importance Analysis: The feature contribution of the random forest model is quantified by the Gini impurity decrease mean. The feature importance of the support vector machine model is ranked by analyzing the contribution of features to the predictive ability. The results of the two models are integrated to identify the key discriminative feature subset. Step Six: 3D Visualization Implementation: Map the brain regions with statistically significant differences from Step Two to the brain regions corresponding to the key features from Step Five onto the standard MNI152 brain template using 3D rendering technology, and generate an interactive 3D visualization atlas using 3D rendering technology.

5. The method for a three-dimensional visualization system of brain regions for AD discrimination based on machine learning and spatial statistics according to claim 4, characterized in that, The formula for calculating the Z-score intensity normalization in step one is as follows: in Original voxel values, and These are the mean and standard deviation of the image pixel values, respectively; the individual brain mask is automatically generated using the NiftiMasker tool in the Nilearn library with a 'whole-brain-template' strategy.

6. The method for a three-dimensional visualization system of brain regions for AD discrimination based on machine learning and spatial statistics as described in claim 4, characterized in that: The predefined regions of interest (ROIs) in step two include 20 brain regions: the white matter of the left and right hemispheres, the cerebral cortex, the lateral ventricles, the thalamus, the caudate nucleus, the putamen, the globus pallidus, the hippocampus, the amygdala, and the nucleus accumbens.

7. The method for a three-dimensional visualization system of brain regions for AD discrimination based on machine learning and spatial statistics according to claim 4, characterized in that, The gray-scale statistical features in step three include nine indicators: mean, median, standard deviation, extreme values, 25th / 75th percentile, skewness, and kurtosis. The morphological and geometric features include seven indicators: volume, number of voxels, surface area, compactness, three-dimensional coordinates of the centroid, and the overlap ratio with the group of horizontal masks. The group mask region features include eight indicators: mean gray value of the group mask region, median gray value of the group mask region, standard deviation of gray value of the group mask region, minimum gray value of the group mask region, maximum gray value of the group mask region, volume of the group mask region, surface area of ​​the group mask region, and number of voxels of the group mask region. The asymmetry features include three indicators: asymmetry in volume between the left and right hemispheres, asymmetry in intensity between the left and right hemispheres, and asymmetry in surface area between the left and right hemispheres. The global features include eight indicators: mean gray level, standard deviation, skewness, kurtosis, sphericity, and equivalent diameter of the ellipsoid for the whole brain and differential brain regions.

8. The method for a three-dimensional visualization system of brain regions for AD discrimination based on machine learning and spatial statistics according to claim 4, characterized in that, The optimized hyperparameters for the random forest in step four include: max_depth=16, max_features=log2, min_samples_leaf=1, min_samples_split=2, n_estimators=476; The optimized hyperparameters of the support vector machine include: C=12.07, gamma=0.053, kernel=rbf; The performance evaluation metrics for the stratified five-fold cross-validation include accuracy, sensitivity for each category, specificity, and macro-average AUC value.

9. The method for a three-dimensional visualization system of brain regions for AD discrimination based on machine learning and spatial statistics according to claim 4, characterized in that, The 3D rendering in step six is ​​implemented using Python's nilearn library and Plotly.js library. The voxel statistics are mapped to the fsaverage5 cortical surface using the surface.vol_to_surf function. The statistics are encoded using RdBu color mapping. The initial statistical threshold is set to |t|>2.

0. Hovering over or clicking on the brain region can display the precise MNI coordinates and statistical values / feature contribution.

10. The method for a three-dimensional visualization system of brain regions for AD discrimination based on machine learning and spatial statistics according to claim 4, characterized in that: In step six, the generation of the 3D visualization map uses the surface.vol_to_surf function to map the statistical parameter map and key features to the fsaverage standard brain surface, extracts 3D geometric data, and creates an interactive 3D mesh object using the Plotly.js library. It uses red and blue to map the statistical significance level or feature contribution and supports multi-view rotation, scaling, hemisphere selection, and interactive querying of brain region coordinates and statistical values.

Images