A Wilson's Disease Auxiliary Classification System Based on Co-variation of Three-Dimensional Surface Morphology and Brain Network
By developing a Wilson's disease auxiliary classification system based on the synergistic changes of three-dimensional surface morphology and brain networks, this system addresses the problems of non-standard selection of regions of interest (ROIs), inconsistent correction strategies, and limited graph theory analysis dimensions in existing technologies. It achieves stable ROI selection, predicted intracranial volume correction, and decision tree-assisted classification, improving the reliability and credibility of clinical classification of Wilson's disease and supporting the automation and clinical application of personalized auxiliary analysis.
Patent Information
- Application Number
- CN202611308577.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-27
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies lack standardized regions of interest (ROI) screening processes in clinical classification assessment and personalized auxiliary analysis of Wilson's disease. Correction strategies are inconsistent during brain network construction, graph theory analysis has a single dimension, and cortico-basal ganglia synergistic change analysis lacks a significance determination mechanism. Furthermore, existing methods are difficult to embed into clinical workflows and cannot provide verifiable discrimination logic.
By establishing a Wilson's disease auxiliary classification system based on the coordinated changes of three-dimensional surface morphology and brain network, a generalized linear model is used to screen stable ROIs, perform estimated intracranial volume correction, combine permutation test to generate significant critical values, and use a pre-trained decision tree rule base for auxiliary classification to generate a structured report.
It improves the reliability and reproducibility of region of interest screening, ensures the accuracy and credibility of graph theory indicators, provides the ability of two-level graph theory indicators to distinguish pathophysiological patterns, enhances the acceptability and credibility of auxiliary classification results, and realizes automated individualized auxiliary analysis.
Smart Images

Figure CN123118518A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of medical image data processing and computer-aided diagnosis technology, and in particular relates to an auxiliary classification system for Wilson's disease based on the synergistic changes of three-dimensional surface morphology and brain network. Background Technology
[0002] Structural magnetic resonance imaging (sMRI), a non-invasive neuroimaging technique, has been widely used in the structural assessment of brain disorders such as Wilson's disease. sMRI-based morphological analysis methods can extract three-dimensional surface morphological indicators such as cortical thickness, surface area, cortical volume, mean curvature, and folding index from various brain regions through cortical reconstruction. Graph theory analysis methods based on these morphological indicators construct covariant networks of brain structures by calculating correlation coefficients between morphological indicators of different brain regions, and then calculate network topological properties such as global efficiency, clustering coefficient, and characteristic path length to assess the information transmission and integration capacity and local aggregation characteristics of the brain network. Furthermore, recent studies have explored the neural mechanisms of motor disorders and psychiatric symptoms in Wilson's disease from the perspective of the "cortical-subcortical circuit" by calculating the volume correlation between cortical regions of interest and basal ganglia nuclei (such as the putamen, globus pallidus, caudate nucleus, and thalamus). These methods can effectively reveal differences in brain structural morphology and network organization patterns between Wilson's disease patients and healthy controls at the population level.
[0003] However, the existing methods mentioned above still have many shortcomings in clinical classification assessment and personalized auxiliary analysis of Wilson's disease. First, the selection of regions of interest lacks a standardized statistical screening process. Existing studies usually include all regions of the whole brain in the analysis based on prior atlases or select brain regions with significant differences based on experience after inter-group comparisons. There is a lack of technical means to screen core regions of interest by combining statistical hypothesis testing with cross-sample stability verification under the premise of controlling for covariates such as age, disease course, and severity of neurological symptoms. This leads to large differences in results between different studies and poor reproducibility. Second, the strategies for adjusting the estimated intracranial volume during brain network construction are inconsistent. Some studies have not performed correction or have only used simple proportional correction, which has failed to effectively eliminate the confounding effect of individual head size differences on the correlation of morphological indicators, directly affecting the reliability of graph theory indicators. Third, the graph theory analysis has a single dimension. Existing methods mostly focus on the overall topological properties of the whole brain network and lack a two-level graph theory indicator joint assessment of the whole brain network and disease-related core sub-networks. It is difficult to distinguish between different pathophysiological patterns of whole-brain network collapse and local compensatory changes. Furthermore, the analysis of cortico-basal ganglia synergistic changes lacks a standardized significance determination mechanism based on permutation tests, resulting in the inclusion of numerous weakly correlated or false-positive connections, interfering with the identification of pathophysiologically significant synergistic change patterns. Simultaneously, the determination of clinical subtypes of Wilson's disease primarily relies on physicians' subjective experience. While a few studies employ deep learning models for classification, these are "black box" models, failing to provide clear and verifiable discriminative logic and exhibiting low clinical acceptability. More critically, existing analytical methods are mostly fragmented tools used in research settings, with each processing step distributed across different software platforms, requiring manual coordination. They fail to consider the need for integration with hospital PACS and HIS systems, do not support individualized data input, structured data output, or the generation of visual reports, and are difficult to embed into clinical workflows for automated and standardized auxiliary analysis. Summary of the Invention
[0004] To address the aforementioned technical problems, this invention proposes an auxiliary classification system for Wilson's disease based on the synergistic changes in three-dimensional surface morphology and brain networks, thereby resolving the issues present in the prior art.
[0005] To achieve the above objectives, in a first aspect, the present invention provides a Wilson's disease auxiliary classification system based on the synergistic changes in three-dimensional surface morphology and brain networks, comprising: The data acquisition module is used to acquire three-dimensional T1-weighted structural magnetic resonance imaging data of the subject's brain; The preprocessing and feature extraction module is used to preprocess the structural magnetic resonance imaging data, reconstruct the surface of the whole brain cortex and divide the brain regions, and extract the three-dimensional surface morphology indicators of each brain region and the volume indicators of the subcortical basal ganglia nuclei. The stable ROI screening module is used to establish a generalized linear model for the three-dimensional surface morphology indicators of each brain region with age, disease course and neurological symptom scores as covariates, perform statistical hypothesis testing, and combine Bootstrap resampling to screen out a set of stable regions of interest across samples. The brain network construction and graph theory calculation module is used to construct a whole-brain morphological correlation matrix after estimating the intracranial volume of the morphological indices of each brain region in the set of regions of interest, and to calculate the graph theory indices of the whole-brain network and the graph theory indices of the sub-networks composed of some of the regions of interest based on the whole-brain morphological correlation matrix. The collaborative change analysis module is used to calculate the volume correlation coefficient between each brain region in the set of regions of interest and the basal ganglia nuclei, and to threshold the volume correlation coefficient based on the significance threshold value generated by the permutation test to generate a binary collaborative change matrix. The classification report module is used to input the graph theory index and the binary co-variance matrix into a pre-trained decision tree rule base, traverse the discrimination logic of the decision tree rule base to output Wilson disease auxiliary classification labels and confidence scores, and generate a report.
[0006] Preferably, the preprocessing includes desquamation, intensity normalization, and spatial normalization; the brain region division uses the Desikan-Killiany-Tourville atlas to divide the cerebral cortex into 68 regions of interest; the volume indicators of the subcortical basal ganglia nuclei include the volumes of the bilateral putamen, globus pallidus, caudate nucleus, and thalamus.
[0007] Preferably, the generalized linear model uses the three-dimensional surface morphology index of a single brain region as the dependent variable and age, disease duration, and neurological symptom scores as covariates, and assesses the significance of differences between groups using the F-test and error detection rate correction; the Bootstrap resampling uses resampling with replacement and calculates the frequency of each brain region being judged as significantly different to screen stable brain regions; the set of stable regions of interest includes the right middle frontal gyrus, left superior frontal gyrus, right precentral gyrus, right insula, left parahippocampal gyrus, and bilateral medial orbitofrontal gyrus.
[0008] Preferably, the correction of the estimated intracranial volume adopts the regression residual method, that is, a linear regression model is established with the estimated intracranial volume as the independent variable and the morphological index as the dependent variable, and the original value of the morphological index is replaced with the residual value of the linear regression model; the whole brain morphological correlation matrix is obtained by calculating the Pearson correlation coefficient between each pair of morphological indices based on the correction.
[0009] Preferably, the graph theory metrics of the whole-brain network include global efficiency, clustering coefficient, and feature path length, and the graph theory metrics of the sub-network include sub-network global efficiency, sub-network clustering coefficient, and sub-network feature path length; the sub-network is composed of all brain regions in the stable set of regions of interest.
[0010] Preferably, the permutation test includes randomly shuffling the sample group labels and performing several repeated samplings to generate a significance threshold for each connection; the thresholding process involves comparing the absolute value of each element in the collaborative change matrix with the corresponding significance threshold to generate the binary collaborative change matrix.
[0011] Preferably, the decision tree rule base is pre-trained using the CART decision tree algorithm with Gini impurity as the splitting criterion; the formula for calculating Gini impurity is: ; Where K is the total number of categories, p k This represents the proportion of samples of class k in node D; The formula for calculating the weighted sum of Gini impurities of child nodes after splitting is: ; Where, N m Let m be the total number of samples in the current node. and These represent the number of samples in the left and right child nodes after the split, respectively. and These are the sample sets for the left and right child nodes, respectively.
[0012] Preferably, the decision tree rule base includes a first global efficiency threshold, a first sub-network efficiency threshold, a second global efficiency threshold, and a second sub-network efficiency threshold; when the global efficiency of the whole brain is lower than the first global efficiency threshold and the global efficiency of the sub-network is higher than the first sub-network efficiency threshold, a first auxiliary classification label is output; when the condition is not met and the global efficiency of the whole brain is lower than the second global efficiency threshold and the global efficiency of the sub-network is lower than the second sub-network efficiency threshold, a second auxiliary classification label is output; otherwise, a third auxiliary classification label is output; the confidence level is the proportion of training samples in the final leaf node of the decision tree when the sample falls into that node.
[0013] Preferably, the report includes a structured data report in JSON format and a visual report in PDF format; the system supports both cloud application programming interface (API) deployment and hospital internal network plugin deployment.
[0014] Secondly, the present invention also discloses an auxiliary classification method for Wilson's disease based on the synergistic changes of three-dimensional surface morphology and brain network, for implementing the system described in the first aspect.
[0015] Compared with the prior art, the present invention has the following advantages and technical effects: This invention utilizes the stable ROI screening module to establish a generalized linear model with age, disease duration, and neurological symptom scores as covariates for statistical hypothesis testing. This effectively eliminates the confounding effects of covariates on the assessment of inter-group differences, avoiding subjective biases in selecting brain regions based on human experience or simple inter-group comparisons. Combined with Bootstrap resampling technology for cross-sample stability verification, this ensures that the selected ROI set exhibits stable inter-group differences across different sample subsets, thereby improving the reliability and reproducibility of the screening results.
[0016] This invention uses the brain network construction and graph theory calculation module to predict and correct the intracranial volume of morphological indicators. It employs the regression residual method to eliminate the confounding effect of individual head size differences on brain region morphological indicators, so that the corrected morphological indicators truly reflect the pathological changes in brain structure rather than individual physiological differences. Based on this, the whole-brain morphological correlation matrix constructed has higher accuracy and population representativeness, avoiding the bias in network edge weight calculation caused by inconsistent correction strategies.
[0017] This invention, through its brain network construction and graph theory computation module, not only calculates the global efficiency, clustering coefficient, and characteristic path length of the whole-brain network, but also simultaneously calculates the corresponding graph theory indices of subnetworks composed of stable regions of interest. The joint evaluation of these two-level graph theory indices can distinguish between two different network alteration patterns: "a general decline in the integrative capacity of the whole-brain network" and "whole-brain inefficiency accompanied by compensatory enhancement of local subnetworks," providing a richer information foundation for explaining the pathophysiological mechanisms of different clinical subtypes of Wilson's disease.
[0018] This invention uses the co-variation analysis module to generate significance thresholds for each connection through permutation tests and performs thresholding on the volume correlation coefficients for each connection. This effectively eliminates false positive correlations caused by limited sample size or data noise, ensuring that all connections retained in the final generated binary co-variation matrix are statistically significant, thereby improving the reliability of identifying abnormal co-variation patterns in the corticobasal ganglia circuit.
[0019] This invention uses the typing report module to input graph theory indicators and binary co-variance matrices into a pre-trained decision tree rule base. It then uses the inherent if-else branch structure of the decision tree to output typing labels and confidence levels. This allows clinicians to clearly trace the specific threshold comparison logic and decision path on which each typing conclusion is based. This overcomes the deficiency of deep learning "black box" models in providing verifiable criteria and improves the acceptability and credibility of auxiliary typing results in clinical decision-making.
[0020] This invention constructs a complete automated processing pipeline from raw MRI data input to classification report output by sequentially coupling a data acquisition module, a preprocessing and feature extraction module, a stable ROI screening module, a brain network construction and graph theory calculation module, a synergistic change analysis module, and a classification report module. The modules are transferred between each other through standardized data objects, eliminating the need for manual connection. The final generated report supports structured data output and visual report output, can interface with hospital PACS and HIS systems, and can be embedded into clinical workflows to achieve individualized auxiliary classification assessment. Attached Figure Description
[0021] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic diagram of an auxiliary classification system for Wilson's disease based on the synergistic changes of three-dimensional surface morphology and brain network, according to an embodiment of the present invention. Figure 2 This is an embodiment of the present invention, showing the architecture and data flow diagram of an auxiliary classification system for Wilson's disease based on the coordinated changes of three-dimensional surface morphology and brain network. Figure 3 This is a schematic diagram of the rule base discrimination logic in an embodiment of the present invention. Detailed Implementation
[0022] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.
[0023] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.
[0024] Example 1 like Figure 1-2As shown, this embodiment provides an auxiliary classification system for Wilson's disease based on the synergistic changes of three-dimensional surface morphology and brain networks, including: The data acquisition module is used to acquire three-dimensional T1-weighted structural magnetic resonance imaging data of the subject's brain; Specifically, it receives T1-weighted structural magnetic resonance imaging data in DICOM or NIfTI format from the hospital's PACS system or uploaded.
[0025] The preprocessing and feature extraction module is used to preprocess the structural magnetic resonance imaging data, reconstruct the surface of the whole brain cortex and divide the brain regions, and extract the three-dimensional surface morphology indicators of each brain region and the volume indicators of the subcortical basal ganglia nuclei. Furthermore, the preprocessing includes desquamation, intensity normalization, and spatial normalization; the brain region division uses the Desikan-Killiany-Tourville atlas to divide the cerebral cortex into 68 regions of interest; the volume indicators of the subcortical basal ganglia nuclei include the volumes of the bilateral putamen, globus pallidus, caudate nucleus, and thalamus.
[0026] Specifically, preprocessing includes: (1) Denoising: using tools such as FSL's BET to dissect the skull and remove non-brain tissue signals. (2) Intensity normalization: using methods such as those proposed by Nyúl et al., the MRI signal intensities obtained from different scanners are mapped to a standard interval to eliminate differences between devices. (3) Spatial normalization: using linear or nonlinear registration algorithms (such as FSL's FLIRT / FNIRT) to register individual brain images to a standard space (such as MNI152), laying the foundation for subsequent population statistical analysis.
[0027] The stable ROI screening module is used to establish a generalized linear model for the three-dimensional surface morphology indicators of each brain region with age, disease course and neurological symptom scores as covariates, perform statistical hypothesis testing, and combine Bootstrap resampling to screen out a set of stable regions of interest across samples. Furthermore, the generalized linear model uses the three-dimensional surface morphology index of a single brain region as the dependent variable and age, disease duration, and neurological symptom scores as covariates. The significance of differences between groups is assessed by F-test and error detection rate correction. The Bootstrap resampling screens stable brain regions by resampling with replacement and calculating the frequency of each brain region being judged as significantly different. The set of stable regions of interest includes the right middle frontal gyrus, left superior frontal gyrus, right precentral gyrus, right insula, left parahippocampal gyrus, and bilateral medial orbitofrontal gyrus.
[0028] This embodiment also includes cortical reconstruction: high-precision reconstruction of the entire brain cortex surface is performed using the recon-all workflow of FreeSurfer software (v7.2 or later). This workflow automatically completes a series of complex operations such as intensity normalization, Talairach transformation, gray and white matter segmentation, and surface topology repair.
[0029] This embodiment also includes standard partitioning: based on the reconstructed cortical surface, the cerebral cortex of each cerebral hemisphere is precisely divided into 34 regions, totaling 68 ROIs, using the Desikan-Killiany-Tourville atlas. This invention preferably uses, but is not limited to, this atlas.
[0030] This embodiment also includes feature extraction: (1) Three-dimensional surface / morphological indicators: Five key morphological indicators are extracted in batches from each of the above ROIs: (2) Cortical thickness: The average distance between the gray matter surface and the white matter surface. (3) Surface area: The area of the ROI on the cortical surface. (4) Cortical volume: The volume calculated from the cortical thickness and surface area. (5) Average curvature: A geometric indicator that describes the degree of cortical folding. (6) Folding index: An indicator that reflects the complexity of local cortical folding. (7) Basal ganglion nucleus volume: At the same time, the volume indicators of subcortical nuclei are extracted from the segmentation results of FreeSurfer, including the bilateral putamen, globus pallidus, caudate nucleus and thalamus.
[0031] Specifically, the process of controlling covariates to screen for stable ROIs under the GLM framework includes: (1) Model establishment: For each candidate ROI, a generalized linear model is established for the morphological indicators (such as cortical thickness). The model uses the morphological indicators as the dependent variable, and the independent variables include age (continuous variable), disease duration (continuous variable, unit: month), and neurological symptom scores (such as N-UWDRS score, continuous variable), and introduces grouping variables (G1 / G2 / G3) to examine the differences between groups.
[0032] (2) Statistical screening: The F-test was used to assess whether the differences between groups of the ROI were statistically significant after controlling for covariates. The significance level was set at p<0.05, and the false discovery rate was corrected to control for multiple comparisons.
[0033] (3) Stability verification: To further ensure that the selected ROIs are not affected by fluctuations in specific samples, the Bootstrap resampling technique (e.g., 1000 samplings) was used to calculate the frequency of each ROI being selected into the significant set. The final set of 7 stable ROIs (right middle frontal gyrus, left superior frontal gyrus, right precentral gyrus, right insula, left parahippocampal gyrus, and bilateral medial orbitofrontal gyrus) are brain regions that have shown excellent performance in both statistical significance and cross-sample stability.
[0034] The brain network construction and graph theory calculation module is used to construct a whole-brain morphological correlation matrix after estimating the intracranial volume of the morphological indices of each brain region in the set of regions of interest, and to calculate the graph theory indices of the whole-brain network and the graph theory indices of the sub-networks composed of some of the regions of interest based on the whole-brain morphological correlation matrix. Furthermore, the correction of the estimated intracranial volume adopts the regression residual method, that is, a linear regression model is established with the estimated intracranial volume as the independent variable and the morphological index as the dependent variable, and the original value of the morphological index is replaced with the residual value of the linear regression model; the whole brain morphological correlation matrix is obtained by calculating the Pearson correlation coefficient between each pair of morphological indices based on the correction.
[0035] Furthermore, the graph theory metrics of the whole-brain network include global efficiency, clustering coefficient, and feature path length, and the graph theory metrics of the sub-network include sub-network global efficiency, sub-network clustering coefficient, and sub-network feature path length; the sub-network is composed of all brain regions in the stable set of regions of interest.
[0036] This embodiment also includes eTIV correction: to eliminate the influence of individual head size on brain structure indicators, a regression residual method is used for correction. That is, a linear regression model is established with eTIV as the independent variable and a certain morphological indicator (such as the volume of a certain ROI) as the dependent variable. The original value of the indicator is replaced with the model residual, which is the standardized indicator after removing the influence of head size.
[0037] This embodiment also includes constructing a correlation matrix: morphological indices (e.g., cortical thickness) of 64 ROIs after eTIV correction for all subjects (n=180) are grouped into 64 variables. Pearson correlation coefficients are calculated between each pair of these variables to form a 64×64 population-level correlation matrix, which characterizes the coordinated change patterns of different brain regions during morphological development or pathological processes.
[0038] Specifically, the calculation of graph theory metrics for the whole brain and ROI subnetworks includes: (1) Graph theory basis: The correlation matrix obtained in the previous step is regarded as the adjacency matrix of the brain structure network, where each node is a brain region and the edge weight is the correlation strength of the morphological indicators of two brain regions.
[0039] (2) Indicator Calculation: The following core graph theory indicators were calculated using the mature Brain Connectivity Toolbox (BCT): (a) Global Efficiency: An indicator of the overall efficiency of network information transmission, calculated as the average of the reciprocals of the shortest path lengths between all node pairs. A higher value indicates faster information transmission. (b) Clustering Coefficient: An indicator of the degree of local clustering in the network, i.e., the degree of interconnection between a node's neighbors. A high value indicates the existence of local functional modules. (c) Feature Path Length: The average of the shortest path lengths between all node pairs in the network. A low value indicates strong network integration capabilities.
[0040] (3) Subnet analysis: In addition to calculating the metrics of the whole brain (64 nodes) network, we also specifically calculated the subnet efficiency, subnet clustering coefficient and subnet characteristic path length of the subnet composed of the 7 selected stable ROIs to reveal the local network attributes that are specifically related to the disease.
[0041] The collaborative change analysis module is used to calculate the volume correlation coefficient between each brain region in the set of regions of interest and the basal ganglia nuclei, and to threshold the volume correlation coefficient based on the significance threshold value generated by the permutation test to generate a binary collaborative change matrix. Furthermore, the permutation test includes randomly shuffling the sample group labels and performing several repeated samplings to generate significance thresholds for each connection; the thresholding process involves comparing the absolute value of each element in the collaborative change matrix with the corresponding significance threshold to generate the binary collaborative change matrix.
[0042] Specifically, calculating and thresholding the volumetric co-variance of “ROI ↔ basal segment” includes: (1) Calculate the correlation matrix: Calculate the correlation between the volume of the 7 stable ROIs and the volume of the 4 basal ganglia nuclei (which can be combined bilaterally) of all subjects to form a volume co-variation matrix of (7 ROIs × 4 nuclei).
[0043] (2) Significance Thresholding: To avoid false positives and interference from weakly correlated connections, the matrix is thresholded. The significance threshold r table is generated through a permutation test: the group labels are randomly shuffled 1000 times, and a random correlation matrix is calculated each time. The absolute value of each matrix element at the 95th quantile is taken as the critical r value for significance at the p<0.05 level. Finally, only strongly correlated connections with absolute values greater than or equal to the corresponding critical r values in the original matrix are retained to generate a binary covariance matrix.
[0044] The classification report module is used to input the graph theory index and the binary co-variance matrix into a pre-trained decision tree rule base, traverse the discrimination logic of the decision tree rule base to output Wilson disease auxiliary classification labels and confidence scores, and generate a report.
[0045] Furthermore, the decision tree rule base includes a first global efficiency threshold, a first sub-network efficiency threshold, a second global efficiency threshold, and a second sub-network efficiency threshold. When the global efficiency of the whole brain is lower than the first global efficiency threshold and the global efficiency of the sub-network is higher than the first sub-network efficiency threshold, a first auxiliary classification label is output. When the condition is not met and the global efficiency of the whole brain is lower than the second global efficiency threshold and the global efficiency of the sub-network is lower than the second sub-network efficiency threshold, a second auxiliary classification label is output. Otherwise, a third auxiliary classification label is output. The confidence level is the proportion of training samples in the final leaf node of the decision tree when the sample falls into that node.
[0046] Furthermore, the report includes a structured data report in JSON format and a visual report in PDF format; the system supports both cloud application programming interface (API) deployment and hospital internal network plugin deployment.
[0047] Specifically, based on the graph theory metrics and co-variance binary matrix of the training set data, the CART decision tree algorithm is used for training to generate executable classification rules. This decision tree recursively divides the data into subsets that are as pure as possible using Gini impurity as the node splitting criterion. The splitting process is achieved by minimizing the following cost function: For any node m Candidate splits on s Its cost function is calculated by splitting the left and right child nodes. The weighted sum of Dr's Gini impurity is used for evaluation: ; in, Nm It is a node m The number of samples, and It represents the number of samples in the left and right child nodes after the split. and These are the sample sets for the left and right child nodes, respectively. Gini impurity. Gini(D) The calculation formula is: Where K is the total number of categories. p k It is a node D The Middle k The proportion of class samples. The algorithm iterates through all features and all possible split points, selecting the splitting scheme that minimizes G(D,s). s* .
[0048] The final decision tree model is solidified into a series of "if-else" logical rules and integrated into the classification and scoring module. For example... Figure 3 As shown, the core of its discrimination logic can be described as follows: T1 (First Global Efficiency Threshold): This is the critical threshold for determining whether the brain's overall network integration capacity is significantly impaired. Its core function is to identify samples where the brain's information transmission efficiency is at its lowest level. In the discrimination logic, it is a prerequisite for triggering the "global inefficiency" mode judgment.
[0049] T2 (First ROI Subnetwork Efficiency Threshold): This is the critical threshold for determining whether there is compensatory enhancement in local key brain regions. Its core function is to further distinguish whether the inefficiency is due to global collapse or accompanied by compensatory responses in local networks when the overall brain efficiency is low (below T1). In the discriminant logic, it is combined with T1 to identify the characteristic G3 pattern of "global inefficiency + local compensation".
[0050] T3 (Second Global Efficiency Threshold) and T4 (Second ROI Subnetwork Efficiency Threshold): These two thresholds are used together to determine whether the whole-brain and local networks exhibit widespread connectivity disruptions. Their core function is to identify samples with low overall brain efficiency and low local subnetwork efficiency. In the discrimination logic, when a sample does not meet the G3 type condition, but its whole-brain efficiency is lower than T3 and its ROI subnetwork efficiency is simultaneously lower than T4, it is judged to exhibit a G2 type pattern of "widespread global-local impairment".
[0051] The decision tree rule base includes: Rule 1: If the conditions "the global efficiency of the whole brain network is lower than the threshold T1" and "the efficiency of the ROI subnetwork is higher than the threshold T2" are met, the system determines that the sample belongs to type G3 (Parkinson's-like). This pattern corresponds to the combined characteristics of "decreased global network integration ability" and "enhanced compensatory function of local key nodes".
[0052] Rule 2: If Rule 1 is not satisfied, but the conditions "global efficiency of the whole brain network is below threshold T3" and "ROI subnetwork efficiency is below threshold T4" are met, then the system determines that the sample belongs to type G2 (dystonia). This pattern corresponds to the characteristic of "widespread impairment of both global and local network connectivity".
[0053] Rule 3: If none of the above rules are met, the system will default to classifying the sample as G1 (without neurological symptoms), characterized by network metrics that are generally better than other subtypes.
[0054] For each decision result, the system simultaneously calculates its confidence level, which is the proportion of training samples in the final leaf node of the decision tree when the sample falls into that node.
[0055] All discrimination thresholds (such as T1, T2, T3, T4) in the rule base are determined through quantile statistics and k-fold cross-validation of the training set data to ensure their generalization ability. These thresholds and rules are fixed as a whole in the system, and an interface is designed to allow safe iteration and updates as the sample size increases in the future.
[0056] Input features include graph theory metrics (E, C, L) for the whole brain and ROI subnets, as well as key connectivity information in the binarized co-transformation matrix. The algorithm automatically generates a decision tree, whose branching rules (e.g., if whole brain efficiency < 0.35 and ROI subnet efficiency > 0.45 then G3) constitute executable and interpretable discriminative logic.
[0057] Output: The system loads this rule base, calculates the features of the new individual, and traverses the decision tree to obtain auxiliary classification labels (G1 / G2 / G3) and confidence scores. Simultaneously, the system compiles a list of key nodes (ROIs with the most significant morphological changes and their direction of change ↑ / ↓), a Top-N table of co-variations (the strongest ROIs and basal ganglia connections), and automatically generates JSON (for machine reading and system integration) and PDF reports (for doctors, including heatmaps, radar charts, textual explanations, and follow-up recommendations).
[0058] This system can be deployed on a local server or in a cloud computing environment. Through a standardized data processing workflow, it enables the automatic generation of auxiliary classification reports from medical imaging data. Specifically, it includes: 1. System input; The system receives structural magnetic resonance imaging (MRI) data or processed individual brain imaging parameters. When the input is raw image data, it is in the DICOM or NIfTI format, which are commonly used in medical imaging. When the input is individual parameter data, the system receives a structured data object containing the following essential information: patient age, disease duration, neurological symptom score, estimated intracranial volume, three-dimensional morphological parameters of each brain region based on standard brain atlas partitions (including but not limited to cortical thickness, surface area, cortical volume, mean curvature, and folding index), and volume data of major basal ganglia nuclei (such as bilateral putamen, globus pallidus, caudate nucleus, and thalamus).
[0059] 2. Data processing flow; After system initialization, the following core steps are executed in a pre-defined standardized process sequence: First, the data access and parsing module verifies and standardizes the input data to ensure the integrity of the data format and content, preparing it for subsequent processing.
[0060] The process then proceeds to the preprocessing and cortical reconstruction module. If the input is raw image data, this module automatically performs a series of preprocessing steps, including image denoising, intensity normalization, and spatial normalization. Then, it uses FreeSurfer software to perform high-precision reconstruction of the entire cerebral cortex surface and divides the cerebral cortex into sixty-four standard regions based on the Desikan-Killiany-Tourville atlas.
[0061] Next, the feature extraction module quantitatively extracts five types of three-dimensional morphological indicators for each brain region and volume data of the basal ganglia nuclei from the above reconstruction results.
[0062] Then, the statistical screening module was activated. Within a generalized linear model framework that controls for covariates such as age, disease duration, and neurological symptom scores, this module selected features for this analysis from a predefined, large-sample-validated set of stable brain regions. This set of stable brain regions included the right middle frontal gyrus, left superior frontal gyrus, right precentral gyrus, right insula, left parahippocampal gyrus, and bilateral medial orbitofrontal gyruses.
[0063] Next, the network construction and graph theory analysis module performs pre-calcium volume correction on the selected brain region morphological indicators and constructs a whole-brain level correlation matrix based on the corrected indicators. This module uses the brain connectome toolbox to calculate the graph theory indicators of the whole-brain network and the sub-networks composed of the aforementioned stable brain regions. The core indicators include global efficiency, clustering coefficient, and feature path length.
[0064] Simultaneously, the co-variation analysis module calculates the correlation matrix between the volumes of stable brain regions and basal ganglia nuclei, and performs thresholding on the correlation matrix based on a pre-determined significance threshold table through permutation tests, thereby obtaining a binary co-variation matrix to identify statistically significant brain region-basal ganglia connections.
[0065] Finally, the subtyping scoring and reporting module integrates all graph theory indicators and co-variation information calculated in the preceding steps, and calls upon a rule base pre-trained and embedded in the system based on a decision tree algorithm for logical discrimination. This rule base contains explicit judgment conditions; for example, when the efficiency of the whole-brain network is below a certain threshold while the efficiency of the subnetwork in the region of interest is above another specific threshold, it tends to be classified as a specific subtype. Based on this, the system outputs auxiliary subtyping labels, corresponding confidence scores, lists key brain regions with significant morphological changes and their directions of change, and summarizes the brain regions with the strongest co-variation-basal ganglia connection pairs.
[0066] 3. System output.
[0067] After the data processing is completed, the system generates two forms of output results.
[0068] The first method is structured data output, using the widely supported JSON format, which clearly includes the typing results, confidence scores, detailed graph theory indicators, a list of key brain regions, and important information on co-variant connections. This format facilitates subsequent automated computer parsing and system integration.
[0069] The second type is a structured report for doctors, using the common PDF format. This report not only includes the classification conclusions and confidence levels, but also visually displays the comparison and synergistic change patterns of graph theory indicators in the form of heatmaps, radar charts, etc., and includes descriptions of key findings and rule-based auxiliary interpretation text.
[0070] 4. System Characteristics The system described in this embodiment effectively solves the technical problem of inconsistent results caused by different processing procedures by standardizing the complex multimodal image data processing and analysis steps into a standardized, automated pipeline. Its output of structured, machine-readable data and interpretable reports provides clinicians with objective and quantitative decision support information.
[0071] Example 2 Based on the same inventive concept, this embodiment also provides an auxiliary classification method for Wilson's disease based on the synergistic changes of three-dimensional surface morphology and brain network, for implementing the system described in Embodiment 1.
[0072] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A Wilson's disease auxiliary classification system based on the synergistic changes of three-dimensional surface morphology and brain network, characterized in that, include: The data acquisition module is used to acquire three-dimensional T1-weighted structural magnetic resonance imaging data of the subject's brain; The preprocessing and feature extraction module is used to preprocess the structural magnetic resonance imaging data, reconstruct the surface of the whole brain cortex and divide the brain regions, and extract the three-dimensional surface morphology indicators of each brain region and the volume indicators of the subcortical basal ganglia nuclei. The stable ROI screening module is used to establish a generalized linear model for the three-dimensional surface morphology indicators of each brain region with age, disease course and neurological symptom scores as covariates, perform statistical hypothesis testing, and combine Bootstrap resampling to screen out a set of stable regions of interest across samples. The brain network construction and graph theory calculation module is used to construct a whole-brain morphological correlation matrix after estimating the intracranial volume of the morphological indices of each brain region in the set of regions of interest, and to calculate the graph theory indices of the whole-brain network and the graph theory indices of the sub-networks composed of some of the regions of interest based on the whole-brain morphological correlation matrix. The collaborative change analysis module is used to calculate the volume correlation coefficient between each brain region in the set of regions of interest and the basal ganglia nuclei, and to threshold the volume correlation coefficient based on the significance threshold value generated by the permutation test to generate a binary collaborative change matrix. The classification report module is used to input the graph theory index and the binary co-variance matrix into a pre-trained decision tree rule base, traverse the discrimination logic of the decision tree rule base to output Wilson disease auxiliary classification labels and confidence scores, and generate a report.
2. The system according to claim 1, characterized in that, The preprocessing includes desquamation, intensity normalization, and spatial normalization; the brain region division uses the Desikan-Killiany-Tourville atlas to divide the cerebral cortex into 68 regions of interest; the volume indicators of the subcortical basal ganglia nuclei include the volumes of the bilateral putamen, globus pallidus, caudate nucleus, and thalamus.
3. The system according to claim 1, characterized in that, The generalized linear model uses three-dimensional surface morphology indices of a single brain region as the dependent variable and age, disease duration, and neurological symptom scores as covariates. The significance of differences between groups is assessed by F-test and error detection rate correction. The Bootstrap resampling uses resampling with replacement and calculates the frequency of each brain region being judged as significantly different to screen stable brain regions. The set of stable regions of interest includes the right middle frontal gyrus, left superior frontal gyrus, right precentral gyrus, right insula, left parahippocampal gyrus, and bilateral medial orbitofrontal gyrus.
4. The system according to claim 1, characterized in that, The correction of the estimated intracranial volume adopts the regression residual method, that is, a linear regression model is established with the estimated intracranial volume as the independent variable and the morphological index as the dependent variable, and the original value of the morphological index is replaced with the residual value of the linear regression model; the whole brain morphological correlation matrix is obtained by calculating the Pearson correlation coefficient between each pair of morphological indices based on the correction.
5. The system according to claim 1, characterized in that, The graph theory metrics of the whole-brain network include global efficiency, clustering coefficient, and feature path length; the graph theory metrics of the sub-network include sub-network global efficiency, sub-network clustering coefficient, and sub-network feature path length; the sub-network is composed of all brain regions in the stable set of regions of interest.
6. The system according to claim 1, characterized in that, The permutation test includes randomly shuffling the sample group labels and performing several repeated samplings to generate significance thresholds for each connection; the thresholding process involves comparing the absolute value of each element in the collaborative change matrix with the corresponding significance threshold to generate the binary collaborative change matrix.
7. The system according to claim 1, characterized in that, The decision tree rule base was pre-trained using the CART decision tree algorithm with Gini impurity as the splitting criterion; the formula for calculating Gini impurity is: ; Where K is the total number of categories, p k This represents the proportion of samples of class k in node D; The formula for calculating the weighted sum of Gini impurities of child nodes after splitting is: ; Where, N m Let m be the total number of samples in the current node. and These represent the number of samples in the left and right child nodes after the split, respectively. and These are the sample sets for the left and right child nodes, respectively.
8. The system according to claim 1, characterized in that, The decision tree rule base includes a first global efficiency threshold, a first sub-network efficiency threshold, a second global efficiency threshold, and a second sub-network efficiency threshold. When the global efficiency of the whole brain is lower than the first global efficiency threshold and the global efficiency of the sub-network is higher than the first sub-network efficiency threshold, a first auxiliary classification label is output. When the condition is not met and the global efficiency of the whole brain is lower than the second global efficiency threshold and the global efficiency of the sub-network is lower than the second sub-network efficiency threshold, a second auxiliary classification label is output. Otherwise, a third auxiliary classification label is output. The confidence level is the proportion of training samples in the final leaf node of the decision tree when the sample falls into that node.
9. The system according to claim 1, characterized in that, The report includes structured data reports in JSON format and visual reports in PDF format; the system supports both cloud application programming interface (API) deployment and hospital internal network plugin deployment.
10. A Wilson's disease auxiliary classification method based on the synergistic changes of three-dimensional surface morphology and brain network, characterized in that, Used to implement the system according to any one of claims 1-9.
