Schizophrenia auxiliary diagnosis and abnormity positioning method and system

Abstract

The invention provides a schizophrenia auxiliary diagnosis and abnormity positioning method and system, and relates to the technical field of artificial intelligence. According to the method, BOLD signal extraction is carried out on each grey matter brain region and white matter brain region in the preprocessed data set, and white matter bold signals are introduced into classification feature construction of schizophrenia for the first time, so that the dimension of brain function analysis is expanded; on this basis, a new TW3C feature is constructed to quantify collaborative activities between grey matter areas under the same white matter pathway, resonance / common collaborative information in the transmission process of grey matter and white matter neural signals is effectively extracted, brain area features with the most discriminative ability are extracted, and diagnosis and research of schizophrenia are facilitated. In addition, based on the classification feature set, an SVM classifier is used for classification, and the brain region corresponding to the feature with the high weight value is positioned.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and system for auxiliary diagnosis and abnormality location of schizophrenia. Background Art

[0002] Schizophrenia is a chronic mental disorder with common clinical symptoms including paranoid delusions, hallucinations, and other perceptual and cognitive abnormalities. Traditional diagnostic methods are time-consuming, limited in timeliness, highly subjective, and lack objective biomarkers. Therefore, there is an urgent need for early diagnosis and precise treatment.

[0003] However, the current diagnostic method for schizophrenia mainly relies on professional psychiatrists to conduct systematic evaluations using a variety of mental scales. Although this method has been widely used in clinical practice, it is inevitably affected by subjective judgment and individual differences during the evaluation process, and the diagnostic results may have certain deviations. In addition, the representation rate of mental illnesses overlaps frequently, and relying on patient self-reports or patient performance status can easily lead to misdiagnosis and missed diagnosis, resulting in delayed intervention for schizophrenia patients. Therefore, exploring more objective and automated diagnostic methods has important research and application value. The auxiliary diagnosis part of this project is aimed at the imaging diagnosis of schizophrenia, and the use of biomarkers for diagnosis is more objective.

[0004] The market demand for smart medical tools is experiencing explosive growth, with disease diagnosis playing a significant role. However, most diagnostic aids currently available on the market focus on underlying conditions, and psychotherapy platforms rely heavily on patient dialogue, making it difficult to provide accurate and explainable diagnostic support.

[0005] Therefore, we need to provide a method and system for auxiliary diagnosis and abnormality localization of schizophrenia to solve the technical problems of inaccurate positioning, incomplete feature extraction, and weak feature correlation in the existing methods for auxiliary diagnosis of schizophrenia. Summary of the Invention

[0006] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a method and system for auxiliary diagnosis and abnormality localization of schizophrenia. It creatively introduces white matter-related features and constructs classification features through white matter BOLD signals, thereby improving the classification accuracy of schizophrenia patients and aiming to achieve the effect of auxiliary diagnosis and abnormality localization.

[0007] To achieve the above objectives, the present application proposes a method for auxiliary diagnosis and abnormality location of schizophrenia, comprising the following steps: Step 1: Obtain resting-state functional magnetic resonance imaging data of each subject and perform data preprocessing one by one to obtain a preprocessed data set; Step 2: Extract BOLD signals from each gray matter region and white matter region in the preprocessed data set to obtain corresponding BOLD signals; Step 3: Combine each BOLD signal at each time point to construct a gray matter time series signal vector and a white matter time series signal vector, and obtain the gray matter time series signal matrix and the white matter time series signal matrix from each vector; Step 4: Based on the gray matter time series signal matrix and the white matter time series signal matrix, calculate the Pearson correlation coefficient of each white matter brain region relative to each gray matter brain region to obtain a Pearson correlation matrix; Step 5: Calculate the TW3C features between specific white matter regions and each gray matter pair based on the Pearson correlation matrix to obtain the TW3C feature sequence corresponding to the specific white matter region. Construct the TW3C feature matrix of all the TW3C feature sequences of the specific white matter regions. Step 6: Quantize the TW3C feature matrix into feature vectors, and assign indexes to its feature vectors. Then, based on the indexes, select the feature vectors to obtain the classification feature set. Step 7: Use the SVM classifier to perform classification based on the classification feature set, and locate abnormalities in the brain areas corresponding to high-weight features.

[0008] As a further solution, the Craddock gray matter template was used to extract the signal average of all voxels in each gray matter brain area to obtain the BOLD signal of each gray matter brain area.

[0009] As a further solution, the ICBM-JHU white matter template was used to extract the signal average of all voxels in each white matter brain region to obtain the BOLD signal of each white matter brain region.

[0010] As a further solution, in step 2, all BOLD signals in the gray matter time series signal matrix and the white matter time series signal matrix are normalized by the z-score normalization method.

[0011] As a further solution, the Pearson correlation coefficient is calculated by the following formula: in, For the Gray matter time series signal vector of gray matter brain regions, is the gray matter time series signal vector The average BOLD signal of For the Gray matter time series signal vector of white matter brain regions, is the white matter time series signal vector The average BOLD signal of T is the total number of time points, t is the time point variable, is the gray matter time series signal vector exist t BOLD signal value at a given time point, is the white matter time series signal vector exist t BOLD signal value at a given time point.

[0012] As a further solution, the TW3C feature is calculated by the following formula: in, Indicates and The corresponding gray matter brain areas form a gray matter pair and are connected by Corresponding specific white matter brain areas The degree of indirect interaction, i1 and i2 Respectively represent the gray matter pair numbers.

[0013] As a further solution, in step 6, random downsampling is performed to select feature vectors; the specific steps include: Perform sampling without replacement from the index sequence consisting of index identifiers to obtain the corresponding sampling sequence; Get the TW3C features in the feature vector corresponding to each index identifier in the sampling sequence; The total number of TW3C features in the TW3C feature matrix is taken as the sum of the upper triangular elements and the matrix dimension is repeatedly derived; The obtained matrix dimensions are used to construct multiple different classification feature subsets.

[0014] As a further solution, the constructed classification feature subset is sorted by feature importance through random forest, and the importance threshold is selected as the classification feature. After feature merging and deduplication, the classification feature set corresponding to the importance threshold is obtained.

[0015] On the other hand, the present invention further provides a schizophrenia auxiliary diagnosis and abnormality localization system, which uses the schizophrenia auxiliary diagnosis and abnormality localization method as described in any one of the above items, comprising: A data preprocessing module is used to obtain the resting-state functional magnetic resonance imaging data of each subject and perform data preprocessing on each subject to obtain a preprocessed data set; The feature extraction module is used to extract the BOLD signal of each gray matter brain region and white matter brain region in the preprocessed data set to obtain the corresponding BOLD signal; Combined with each time point, each BOLD signal is constructed into a gray matter time series signal vector and a white matter time series signal vector, and the gray matter time series signal matrix and the white matter time series signal matrix are obtained from each vector; A feature construction module is used to calculate the Pearson correlation coefficient of each white matter brain region relative to each gray matter brain region based on the gray matter time series signal matrix and the white matter time series signal matrix to obtain a Pearson correlation matrix; Based on the Pearson correlation matrix, the TW3C features between specific white matter brain regions and each gray matter pair were calculated to obtain the TW3C feature sequence corresponding to the specific white matter brain region. The TW3C feature sequences of all specific white matter brain regions were constructed into a TW3C feature matrix. The feature classification module is used to vectorize the TW3C feature matrix, index its feature vectors, and then select the feature vectors based on the index to obtain the classification feature set; The abnormality localization module uses the SVM classifier to perform classification based on the classification feature set and locates abnormalities in the brain areas corresponding to high-weight features.

[0016] Compared with related technologies, the method and system for auxiliary diagnosis and abnormality localization of schizophrenia provided by the present invention have the following advantages: 1. This paper extracts BOLD signals from each gray matter and white matter region in the preprocessed data set, and for the first time introduces white matter bold signals into the classification feature construction of schizophrenia, expanding the dimension of brain function analysis; 2. This paper constructs a new TW3C feature to quantify the collaborative activity between gray matter regions in the same white matter pathway, effectively extracting the most discriminative brain region features during the transmission of gray and white matter neural signals. This is helpful for the diagnosis and research of schizophrenia. 3 The present invention also uses an SVM classifier to perform classification based on the classification feature set and locates the brain area corresponding to the features with higher weights. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0019] Figure 1A schematic diagram of the steps of a method for auxiliary diagnosis and abnormality location of schizophrenia provided by the present invention; Figure 2 A schematic diagram of the structure of a schizophrenia auxiliary diagnosis and abnormality location system provided by the present invention; The purpose, features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0021] Before we delve into specifics, it's important to understand that in recent years, functional magnetic resonance imaging (fMRI), with its advantages of non-invasiveness and high spatiotemporal resolution, has played a significant role in the study of the pathological mechanisms of schizophrenia (SZ). With the advancement of neuroimaging, numerous medical research centers worldwide have begun independently collecting brain imaging data from patients with schizophrenia, which has been widely used in research on disease mechanisms and as an aid to diagnosis. Furthermore, several international research institutions have released fMRI-based schizophrenia neuroimaging datasets, such as the COBRE and UCLA datasets, providing a rich data foundation for computer-assisted diagnosis. Leveraging these fMRI data, researchers have made significant progress in computational classification methods for SZ. In particular, the widespread application of traditional machine learning and deep learning techniques in this field has significantly improved the accuracy of SZ imaging analysis and aided diagnosis. Computational models not only effectively identify complex, high-dimensional patterns in fMRI data but also provide new research directions for exploring the underlying neurobiological mechanisms of SZ.

[0022] It is worth noting that although deep learning has stronger feature extraction capabilities in theory, its performance in SZ classification tasks is not always better than traditional machine learning methods. Moreover, in recent years, more research has been based on improvements to traditional machine learning classification methods rather than deep learning classification algorithms that are more popular in theory. The reasons for this may be the following two points: First, deep learning generally has a "black box" problem, and the model decision-making process is difficult to explain, which limits its transparency and clinical trust in medical diagnosis. Second, the amount of fMRI data is relatively limited. Due to factors such as high acquisition costs, cross-center data usage rights, and differences in scanning instruments, it is difficult to meet the needs of deep learning for large-scale training samples, resulting in insufficient model training and limited performance. Based on the above reasons, this application chooses to use a classification method based on traditional machine learning to carry out classification research on schizophrenia.

[0023] Example 1 See also Figure 1 This embodiment provides a method for auxiliary diagnosis and abnormality location of schizophrenia, including the following steps: Step 1: Obtain resting-state functional magnetic resonance imaging data of each subject and perform data preprocessing one by one to obtain a preprocessed data set; Step 2: Extract BOLD signals from each gray matter region and white matter region in the preprocessed data set to obtain the corresponding BOLD signals; Step 3: Combine each BOLD signal at each time point to construct a gray matter time series signal vector and a white matter time series signal vector, and obtain the gray matter time series signal matrix and the white matter time series signal matrix from each vector; Step 4: Based on the gray matter time series signal matrix and the white matter time series signal matrix, calculate the Pearson correlation coefficient of each white matter brain region relative to each gray matter brain region to obtain a Pearson correlation matrix; Step 5: Calculate the TW3C features between specific white matter regions and each gray matter pair based on the Pearson correlation matrix to obtain the TW3C feature sequence corresponding to the specific white matter region. Construct the TW3C feature matrix of all the TW3C feature sequences of the specific white matter regions. Step 6: Quantize the TW3C feature matrix into feature vectors, and assign indexes to its feature vectors. Then, based on the indexes, select the feature vectors to obtain the classification feature set. Step 7: Use the SVM classifier to perform classification based on the classification feature set, and locate abnormalities in the brain areas corresponding to high-weight features.

[0024] It's important to note that schizophrenia is a complex mental illness characterized by complex structural and functional connectivity abnormalities. Traditional imaging diagnostic analysis methods primarily focus on gray matter structure, ignoring multidimensional information such as white matter and functional connectivity, making it difficult to fully reveal the disease's mechanisms.

[0025] Therefore, this embodiment creatively introduces white matter-related features and constructs classification features based on white matter BOLD signals, thereby improving the classification accuracy of schizophrenia patients and achieving the effects of auxiliary diagnosis and abnormality localization. To this end, this embodiment extracts BOLD signals from gray matter and white matter brain regions respectively, and constructs a three-way cross-correlation feature (TW3C feature) that integrates white matter and gray matter information of brain regions to distinguish schizophrenia patients from healthy controls.

[0026] In a specific example, we used two multi-center fMRI datasets from public platforms: COBRE (Center for Biomedical Research Excellence): This dataset contains raw structural and functional MRI data from 72 patients with schizophrenia and 75 healthy controls. All subjects were diagnosed using a structured clinical interview based on the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5). Subjects with neurological disorders, mental retardation, a history of severe head trauma (loss of consciousness exceeding 5 minutes), or a history of substance abuse or dependence within the past 12 months were excluded prior to enrollment. Imaging data were acquired using a 3T Siemens scanner, with all subjects resting.

[0027] UCLA (Consortium for Neuropsychiatric Phenomics LA5c Study): The original dataset included 50 SZ patients and 130 healthy controls. This study ultimately enrolled 43 SZ patients and 98 healthy controls. Exclusion criteria included missing T1-weighted structural or rs-fMRI scans, as well as pregnancy, history of head injury, or other contraindications for scanning. Imaging data were also acquired using a 3T Siemens scanner, with all subjects resting.

[0028] Resting-state functional MRI data were preprocessed using the DPABI toolbox. The main preprocessing steps included: 1) removing the first 10 time points to eliminate the effects of initial signal instability. The original time point for the COBRE data was 150, and the original time point for the UCLA data was 152; 2) temporal slice correction, using the middle slice as a reference for slice timing adjustment; 3) motion correction, removing subjects with head movement greater than 2 mm or rotation greater than 2°; 4) scalp structure removal to reduce the influence of irrelevant structures on registration; and noise removal, such as cerebrospinal fluid noise; 5) spatial registration, registering the structural and functional images to the Montreal Neurological Institute (MNI) space; 6) bandpass filtering (0.01–0.1 Hz) to remove high- and low-frequency noise; and 7) normalization and spatial smoothing to ensure spatial consistency.

[0029] After preprocessing, 20 subjects were excluded from the COBRE dataset (15 SZ subjects and 5 HC subjects), while no subjects were excluded from the UCLA dataset. Ultimately, the COBRE dataset included imaging data from 57 SZ subjects and 69 HC subjects, while the UCLA dataset included imaging data from 43 SZ subjects and 98 HC subjects. Detailed demographic information is shown in Table 1.

[0030] Table 1 Demographic information For the preprocessed fMRI data, the present invention uses the Craddock gray matter template and ICBMJHU white matter template, specifically the Craddock-200 gray matter template and the ICBMJHU-48 white matter template (templates of other specifications can also be used) to extract the average BOLD signals of the gray matter and white matter brain regions of all subjects. The specific process is as follows: BOLD signal extraction of gray matter areas: For each gray matter area of each subject , and its corresponding BOLD signal is the average of the signals of all voxels in the area.

[0031] For each time point , No. The BOLD signal of a gray matter area can be expressed as: in, Indicates a time point Next The BOLD signal of the voxel, It is The total number of pixels in the area covered by the gray matter template.

[0032] The fMRI data of each subject are subjected to signal feature extraction using the Craddock200 gray matter template to obtain a Grey Matter Matrix of BOLD time series, where is the number of time points, and 200 is the number of gray matter brain regions.

[0033] BOLD signal extraction of white matter regions: Similarly, for each white matter region , the corresponding BOLD signal is the average of the signals of all voxels in the region. , No. The BOLD signal of each white matter region can be expressed as: in, Indicates a time point Next The BOLD signal of the voxel, It is The total number of pixels in the area covered by the white matter template.

[0034] After extracting the signal features of each subject’s fMRI data using the ICBM JHU48 white matter template, a White Matter Matrix of BOLD time series, where is the number of time points, and 48 is the number of white matter brain regions.

[0035] To avoid the impact of different feature scales on classification performance, the z-score normalization method is used to normalize all signals in the matrix. Specifically, for each signal matrix, its mean μ and standard deviation σ are calculated, and the original signal value x is normalized according to the following formula: The standardized signal has zero mean and unit standard deviation, ensuring that all features participate in model training at the same scale, improving the convergence efficiency and classification stability of the model.

[0036] The above two time series matrices serve as input for subsequent feature construction.

[0037] When constructing features, we aim to capture the co-activation relationships between gray matter brain regions that are indirectly transmitted through white matter pathways. The specific method is as follows: For each subject, the Pearson correlation coefficient of each white matter brain region relative to each gray matter brain region was first calculated. r , as shown in the following formula: in, For the Gray matter time series signal vector of gray matter brain regions, is the gray matter time series signal vector The average BOLD signal of For the Gray matter time series signal vector of white matter brain regions, is the white matter time series signal vector The average BOLD signal of T is the total number of time points, t is the time point variable, is the gray matter time series signal vector exist t BOLD signal value at a given time point, is the white matter time series signal vector exist t BOLD signal value at a given time point.

[0038] For each pair of GM-WM brain regions, the functional connectivity between them was calculated and extracted.

[0039] Each subject will generate a dimension The Pearson correlation matrix (person correlation matrix).

[0040] The matrix contains 9,600 PC values, and each PC value ranges from [-1, 1].

[0041] Among them, the larger the absolute value of the PC value, the stronger the correlation between brain regions; and the closer the absolute value of the PC value is to 0, the weaker the correlation.

[0042] For any pair of gray matter and , which passes through a white matter tract (such as Corresponding specific white matter brain areas The degree of indirect interaction between the gray matter region and the white matter region is defined as the product of its correlation coefficient with the white matter region, namely the Triple Ways Cross Correlation Coefficient (TW3C, as shown in the formula below). This coefficient is used to measure the synchronization of signal transmission from different gray matter regions through the same white matter region.

[0043] in, , Indicates and The corresponding gray matter brain areas form a gray matter pair and are connected by Corresponding specific white matter brain areas The degree of indirect interaction, i1 and i2 Respectively represent the gray matter pair numbers.

[0044] It is noteworthy that each white matter tract (e.g. Corresponding specific white matter brain areas ), there will be corresponding (19,900) TW3C features , where 200 is the number of gray matter template brain regions.

[0045] Finally, the feature sequence extracted for each subject will contain (955200) TW3C feature values, 48 is the number of white matter template brain regions. In addition, the features are vectorized so that each feature has a unique identifier - index , which facilitates the subsequent mining of the most contributing features, as shown in the formula: in, express Corresponding specific white matter brain areas The TW3C characteristic sequence, represents the TW3C feature matrix, vec Indicates that the TW3C matrix is expanded into a one-dimensional vector by row.

[0046] Furthermore, during feature selection, because the number of features constructed in this study far exceeds the number of subjects, significant overfitting can occur. To avoid this problem, this study combined random downsampling and random forests to filter the constructed features. This approach not only reduced the total amount of feature data, speeding up data processing efficiency, but also reduced the risk of overfitting.

[0047] First, the sampling sequence is obtained by sampling without replacement from the index sequence [0, 955200] From each subject's TW3C feature vector Extract Corresponding position Eigenvalue.

[0048] The idea behind selecting the downsampling dimension is as follows: the total feature number 955200 is regarded as the sum of the upper triangular elements of a matrix, and the matrix dimension is derived from this. If the calculated value is not an integer, it is rounded up.

[0049] As shown in the following formula, the matrix dimension is 1383.

[0050] Then, the above process is repeated multiple times to construct multiple different classification feature subsets. This multiple random sampling method avoids the impact of single randomness on the model experimental results and reduces the model's dependence on specific data.

[0051] In this study, the number of random sampling is set to 1000. After 1000 samplings, a relatively balanced data distribution can be obtained. Therefore, 1000 sets of classification feature subsets will be obtained in this stage, which can be expressed as .

[0052] Finally, the classification feature subset is subjected to random forest to obtain the feature importance ranking, and the relative importance threshold is selected as The features of the 1000 samples are used as classification features for mutual verification and comparison. The features obtained after merging and deduplication 1000 times are used as the final classification feature set under this threshold. Therefore, four classification feature subsets are obtained in this stage.

[0053] After completing the construction and screening of TW3C features, the present invention inputs the final four sets of classification feature subsets into the Support Vector Machine (SVM) classifier to evaluate its classification performance in distinguishing schizophrenia patients from healthy controls.

[0054] Generally speaking, features with higher importance weights or features that appear more frequently are considered to be the most contributing brain regions or functional networks. However, it should be noted that due to the randomness of the random forest algorithm itself, the importance weight alone cannot fully determine the contribution of a feature.

[0055] To this end, we counted the frequency of occurrence of each feature in the classification feature set based on its index and sorted them by frequency, determining the most frequently occurring feature as the most contributing classification feature. The higher the frequency of a feature, the more important it is in the random forest model under different sampling conditions.

[0056] Furthermore, because the TW3C feature value represents the cross-correlation between "GM-WM-GM," the TW3C in the classification feature set can be split into two different gray matter and one white matter. The frequency of TW3C is then counted, and the brain region with the most occurrences is considered the most contributing brain region. A more frequent occurrence of a brain region indicates its participation in multiple important connected networks. Although the feature subsets selected under different datasets and different thresholds vary, brain regions closely associated with schizophrenia are consistently identified.

[0057] Example 2 See also Figure 2 Based on Example 1, this embodiment provides a schizophrenia auxiliary diagnosis and abnormality location system, including: A data preprocessing module is used to obtain the resting-state functional magnetic resonance imaging data of each subject and perform data preprocessing on each subject to obtain a preprocessed data set; The feature extraction module is used to extract the BOLD signal of each gray matter brain region and white matter brain region in the preprocessed data set to obtain the corresponding BOLD signal; Combined with each time point, each BOLD signal is constructed into a gray matter time series signal vector and a white matter time series signal vector, and the gray matter time series signal matrix and the white matter time series signal matrix are obtained from each vector; A feature construction module is used to calculate the Pearson correlation coefficient of each white matter brain region relative to each gray matter brain region based on the gray matter time series signal matrix and the white matter time series signal matrix to obtain a Pearson correlation matrix; Based on the Pearson correlation matrix, the TW3C features between specific white matter brain regions and each gray matter pair were calculated to obtain the TW3C feature sequence corresponding to the specific white matter brain region. The TW3C feature sequences of all specific white matter brain regions were constructed into a TW3C feature matrix. The feature classification module is used to vectorize the TW3C feature matrix, index its feature vectors, and then select the feature vectors based on the index to obtain the classification feature set; The abnormality localization module uses the SVM classifier to perform classification based on the classification feature set and locates abnormalities in the brain areas corresponding to high-weight features.

[0058] It should be noted that: in the specific implementation process, the feature construction module first calculates the Pearson correlation coefficient between the gray matter and white matter feature information extracted in the previous step, and generates a value of The correlation coefficient matrix is generated. Next, the TW3C eigenvalues between a specific white matter and gray matter pair are calculated and all TW3C eigenvalues are combined into a one-dimensional feature vector. Subsequently, the one-dimensional feature vector generated in the previous step is randomly downsampled and features are selected based on importance and thresholds to form the final classification feature subset. Finally, the feature classification module uses a support vector machine (SVM) classifier based on the classification feature set and locates the brain regions corresponding to features with higher weights.

[0059] The method proposed in this study effectively extracts the resonance / cooperation information in the transmission process of gray matter and white matter neural signals and extracts the most discriminative brain region features, which is helpful for the diagnosis and research of schizophrenia.

[0060] In a specific experiment, the experimental data are shown in Tables 2 and 3 below: Table 2 Comparison with other methods on the cobre dataset Table 3 Comparison with other methods on the UCLA dataset Song: This method constructs a weighted brain hypernetwork represented by an adjacency tensor. A new hyperedge weight estimation method is proposed to convert the association matrix into a weighted adjacency tensor. Hypergraph signal processing methods such as hypergraph Fourier transform and spectral analysis are applied to perform classification using hypergraph spectra and spectral signals.

[0061] Wang: This method uses Pearson and Spearman rank correlation to construct a sparse matrix based on SRC, and proposes a multi-sparse connectivity network (MSCN) based on deep learning. Gallos: Constructs a functional connectivity network (FN) by correlating packet time series and obtains a low-dimensional embedding of the correlation matrix using Isomap. Five local graph-theoretic features are extracted from the FCN. Feature selection is performed using Lasso and random forest, followed by classification using a linear support vector machine. Wang: This method introduces a multi-kernel capsule network (mkcapsnet) that integrates feature extraction and classification into a unified framework. Multiple kernels are designed based on brain anatomical segmentation to capture functional connectivity at different spatial scales. The capsule structure enables joint optimization of features and classifiers, and capsule dropout is used to prevent overfitting. This paper introduces a new approach, called Multi-Level Representation with Multimodal Imaging and Multi-Classifier (M3), to classify schizophrenia (SZ) and healthy controls (HC) using fMRI and sMRI data. Zhu: This method introduces Temporal-BCGCN, which consists of three parts: DSF-BrainNet for dynamic synchronization feature extraction, TemporalConv for graph convolution based on temporal attributes, and CategoryPool for hemispheric lateralization analysis using RS-fMRI data. Tavakoli: This method uses structural MRI and resting-state fMRI data to extract volumetric and cortical thickness features, as well as four graph-based functional measures: degree, betweenness centrality, participation coefficient, and local efficiency. Traditional machine learning algorithms were used to classify schizophrenia patients and their subtypes. To reduce dimensionality, feature selection methods were applied, including the minimum redundancy maximum association method, which identifies features that are both informative and minimally redundant. sunil: This method applies deep graph convolutional neural networks and traditional machine learning models to resting-state fMRI data, using 69 graph-theoretic features computed from functional connectivity between brain regions. A graph neural network architecture is employed to jointly learn brain network structure and node features for schizophrenia detection.

[0062] The experimental results of this paper show that the proposed TW3C feature, which measures the degree of synchronization of signals transmitted along the same white matter pathway between different gray matter regions, not only significantly improves the accuracy of SZ classification (reaching 98.53% and 98.19% on the COBRE and UCLA datasets, respectively); this result not only verifies the effectiveness of white matter features in SZ diagnosis, but also further proves the scientific nature and rationality of this research method.

[0063] The above are only some embodiments of the present application and are not intended to limit the patent scope of the present application. All equivalent structural transformations made using the contents of the present application specification and drawings under the technical concept of the present application, or direct / indirect application in other related technical fields are included in the patent protection scope of the present application.

Claims

1. A method for auxiliary diagnosis and abnormal location of schizophrenia, characterized in that: The steps include: Step 1: Obtain resting-state functional magnetic resonance imaging data of each subject and perform data preprocessing one by one to obtain a preprocessed data set; Step 2: Extract BOLD signals from each gray matter region and white matter region in the preprocessed data set to obtain corresponding BOLD signals; Step 3: Combine each BOLD signal at each time point to construct a gray matter time series signal vector and a white matter time series signal vector, and obtain the gray matter time series signal matrix and the white matter time series signal matrix from each vector; Step 4: Based on the gray matter time series signal matrix and the white matter time series signal matrix, calculate the Pearson correlation coefficient of each white matter brain region relative to each gray matter brain region to obtain a Pearson correlation matrix; Step 5: Calculate the TW3C features between specific white matter regions and each gray matter pair based on the Pearson correlation matrix to obtain the TW3C feature sequence corresponding to the specific white matter region. Construct the TW3C feature matrix of all the TW3C feature sequences of the specific white matter regions. Step 6: Quantize the TW3C feature matrix into feature vectors, and assign indexes to its feature vectors. Then, based on the indexes, select the feature vectors to obtain the classification feature set. Step 7: Use the SVM classifier to perform classification based on the classification feature set, and locate abnormalities in the brain areas corresponding to high-weight features.

2. The method for auxiliary diagnosis and abnormality location of schizophrenia according to claim 1, characterized in that: The Craddock gray matter template was used to extract the signal average of all voxels in each gray matter brain area to obtain the BOLD signal of each gray matter brain area.

3. The method for auxiliary diagnosis and abnormality location of schizophrenia according to claim 1, characterized in that: The BOLD signal of each white matter region was obtained by extracting the signal average of all voxels in each white matter region using the ICBM-JHU white matter template.

4. The method for auxiliary diagnosis and abnormality location of schizophrenia according to claim 1, characterized in that: In step 2, all BOLD signals in the gray matter time series signal matrix and the white matter time series signal matrix were normalized by the z-score normalization method.

5. The method for auxiliary diagnosis and abnormality location of schizophrenia according to claim 1, characterized in that: The Pearson correlation coefficient is calculated using the following formula: in, For the Gray matter time series signal vector of gray matter brain regions, is the gray matter time series signal vector The average BOLD signal of For the Gray matter time series signal vector of white matter brain regions, is the white matter time series signal vector The average BOLD signal of T is the total number of time points, t is the time point variable, is the gray matter time series signal vector exist t BOLD signal value at a given time point, is the white matter time series signal vector exist t BOLD signal value at a given time point.

6. The method for auxiliary diagnosis and abnormality location of schizophrenia according to claim 5, characterized in that: The TW3C feature is calculated using the following formula: in, Indicates and The corresponding gray matter brain areas form a gray matter pair and are connected by Corresponding specific white matter brain areas The degree of indirect interaction, i1 and i2 Respectively represent the gray matter pair numbers.

7. The method for auxiliary diagnosis and abnormality location of schizophrenia according to claim 1, characterized in that: In step 6, random downsampling is performed to select feature vectors; the specific steps include: Perform sampling without replacement from the index sequence consisting of index identifiers to obtain the corresponding sampling sequence; Get the TW3C features in the feature vector corresponding to each index identifier in the sampling sequence; The total number of TW3C features in the TW3C feature matrix is taken as the sum of the upper triangular elements and the matrix dimension is repeatedly derived; The obtained matrix dimensions are used to construct multiple different classification feature subsets.

8. The method for auxiliary diagnosis and abnormality location of schizophrenia according to claim 7, characterized in that: The constructed classification feature subsets are sorted by feature importance through random forest, and the importance threshold is selected as the classification feature. After feature merging and deduplication, the classification feature set corresponding to the importance threshold is obtained.

9. A schizophrenia auxiliary diagnosis and abnormality localization system, using a schizophrenia auxiliary diagnosis and abnormality localization method according to any one of claims 1 to 8, characterized in that: include: A data preprocessing module is used to obtain the resting-state functional magnetic resonance imaging data of each subject and perform data preprocessing on each subject to obtain a preprocessed data set; The feature extraction module is used to extract the BOLD signal of each gray matter brain region and white matter brain region in the preprocessed data set to obtain the corresponding BOLD signal; Combined with each time point, each BOLD signal is constructed into a gray matter time series signal vector and a white matter time series signal vector, and the gray matter time series signal matrix and the white matter time series signal matrix are obtained from each vector; A feature construction module is used to calculate the Pearson correlation coefficient of each white matter brain region relative to each gray matter brain region based on the gray matter time series signal matrix and the white matter time series signal matrix to obtain a Pearson correlation matrix; Based on the Pearson correlation matrix, the TW3C features between specific white matter brain regions and each gray matter pair were calculated to obtain the TW3C feature sequence corresponding to the specific white matter brain region. The TW3C feature sequences of all specific white matter brain regions were constructed into a TW3C feature matrix. The feature classification module is used to vectorize the TW3C feature matrix, index its feature vectors, and then select the feature vectors based on the index to obtain the classification feature set; The abnormality localization module uses the SVM classifier to perform classification based on the classification feature set and locates abnormalities in the brain areas corresponding to high-weight features.

Images