Auxiliary diagnostic device for bipolar disorder

Abstract

This invention discloses an auxiliary diagnostic device for bipolar disorder, comprising: a structural image preprocessing unit, a functional image preprocessing unit, a structural image feature extraction unit, a functional image feature extraction unit, a feature fusion unit, a classifier training unit, and an auxiliary diagnostic output unit. This invention combines multimodal information from the patient's magnetic resonance structural and functional images to fuse features such as morphology, low-frequency amplitude signals, voxel symmetry homotopy connectivity signals, functional network attributes, and functional network edges, fully utilizing existing information to improve the accuracy of the auxiliary diagnostic model. Simultaneously, to address the overfitting phenomenon caused by limited datasets and a large number of features in medical images, PCA is used for feature dimensionality reduction, and filtering is employed for feature selection to reduce feature redundancy. This effectively selects features with significant differences for classification and diagnosis, further improving the accuracy of the auxiliary diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of machine learning-based auxiliary diagnostic technology, specifically relating to an auxiliary diagnostic device for bipolar disorder. Background Technology

[0002] Bipolar disorder (BD) is a mental illness characterized by periods of manic and depressive moods, also known as mood bipolar disorder. It is one of the earliest onset mental illnesses among modern major mental illnesses and is incurable. Unlike the mood swings experienced by everyone, BD patients experience more pronounced mood fluctuations. During manic episodes, their mood is elevated, with increased interest and motivation; conversely, during depressive episodes, their mood is low, with decreased interest and fatigue, and they may experience suicidal ideation and behavior. Therefore, the disease is characterized by high incidence, high disability rate, and high suicide rate. However, to date, there is still no effective method for diagnosing bipolar disorder. A diagnosis is made by doctors through interviews with the patient and their family. Doctors compare the interview results with diagnostic criteria, commonly using the Diagnostic and Statistical Manual of Mental Disorders (DSM), to determine whether the patient has the disease and the severity of their condition. This process is highly subjective, and finding physiological markers for diagnosis remains a goal for many researchers. Summary of the Invention

[0003] This invention provides an auxiliary diagnostic device for bipolar disorder. It uses machine learning methods to analyze multimodal magnetic resonance imaging data of bipolar disorder patients and uses machine learning to assist in the diagnosis of bipolar disorder, thereby improving the accuracy of diagnosis of bipolar disorder.

[0004] The auxiliary diagnostic device for bipolar disorder provided by the present invention includes: a structural image preprocessing unit, a functional image preprocessing unit, a structural image feature extraction unit, a functional image feature extraction unit, a feature fusion unit, a classifier training unit, and an auxiliary diagnostic output unit;

[0005] The structural image preprocessing unit is used to preprocess the input magnetic resonance structural image, including: segmenting the image according to different tissue structures, and outputting the segmented gray matter image to the structural image feature extraction unit;

[0006] The functional image preprocessing unit is used to preprocess the input MRI functional images, including: removing a specified number of time points at the start of the MRI functional image scan, arranging the corresponding scan layers in chronological order based on the machine's scan sequence, performing correction processing using a preset head motion correction method, establishing the correspondence between the spatial positions of functional images and structural images, and performing spatial standardization, spatial smoothing, delinear drift removal, and signal filtering; and inputting the preprocessed MRI functional image sequence into the functional image feature extraction unit.

[0007] The structural image feature extraction unit resizes the input gray matter image to match the resolution of a preset first brain region partitioning template. It then extracts gray matter from the image according to the template, resulting in image blocks for each brain region. The unit measures the gray matter volume of each region and calculates the Pearson correlation coefficient between any two regions. Based on a preset threshold, it performs binary classification on the Pearson correlation coefficient, assigning a value of 1 to strong correlations and 0 to weak correlations. An adjacency matrix is ​​constructed based on the assigned Pearson correlation coefficients for all brain regions. Each brain region is treated as a node, and a weighted undirected brain structure network is built based on this adjacency matrix. Finally, based on specified network metrics and global network attribute information, the unit extracts the feature information of the brain structure network, forming structural image features. After dimensionality reduction, these features are input into the feature fusion unit.

[0008] The functional image feature extraction unit includes a brain functional network feature extraction unit, a low-frequency amplitude signal feature extraction unit, and a voxel symmetric homotopy connectivity feature extraction unit;

[0009] The brain functional network feature extraction unit is used to resize each image of the input MRI functional image sequence, ensuring the resized image size matches the resolution of a preset second brain region partitioning template. It then partitions each image of the MRI functional image sequence using the second brain region partitioning template, obtaining a time series for each brain region. The unit calculates the correlation coefficient between any two brain region time series using partial correlation coefficients, and binarizes the correlation coefficients using a specified density threshold. If the correlation coefficient is greater than or equal to the density threshold, it is binarized to 1; otherwise, it is binarized to 0. An adjacency matrix of brain regions is constructed based on the binarized correlation coefficients, treating each brain region as a node. A weighted undirected brain functional network is then constructed based on this adjacency matrix. Finally, the unit extracts feature information from the brain functional network based on specified network indicators, filters the extracted brain functional network features for feature selection, and inputs the selected brain functional network features into the feature fusion unit.

[0010] The low-frequency amplitude signal feature extraction unit is used to extract the low-frequency amplitude signal features of the magnetic resonance functional image sequence, and then input the feature fusion unit after feature selection by filtering method.

[0011] The voxel symmetric homotopy connectivity feature extraction unit is used to extract voxel symmetric homotopy connectivity features from magnetic resonance functional imaging sequences, and the features are selected by filtering before being input into the feature fusion unit.

[0012] The feature fusion unit fuses the input features to obtain the multi-modal features of the current object; if the current object is a training object, its multi-modal features are input to the classifier training unit; if the current object is an object to be assisted in diagnosis, its multi-modal features are input to the auxiliary diagnosis output unit.

[0013] The classifier training unit constructs a training dataset based on the multimodal features of the training objects: It sets a classification label for bipolar disorder for each training object based on whether it belongs to the normal group or the disease group. Based on the multimodal features and classification labels of a certain number of training objects, it trains multiple different classifiers until a preset training termination condition is met. Then, it outputs the trained classifications to the auxiliary diagnostic output unit. Each classifier is used to output the predicted binary classification result.

[0014] The auxiliary diagnosis output unit inputs the multi-modal features of the object to be diagnosed into each classifier of this unit, obtains the prediction results of each classifier based on its output, and uses a voting method to determine the diagnosis result of the current object to be diagnosed and outputs it.

[0015] Furthermore, when extracting brain functional network features, the brain functional network feature extraction unit extracts the topological attribute features of the brain functional network and extracts the specified network features of the network corresponding to the upper triangular matrix of the adjacency matrix of the brain region. However, when extracting network features of the network corresponding to the upper triangular matrix of the adjacency matrix of the brain region, the topological attribute features of the network are not included.

[0016] Furthermore, the structural image feature extraction unit uses principal component analysis (PCA) to reduce the dimensionality of the structural image features.

[0017] Furthermore, the classifiers used in the classifier training unit include three types: Support Vector Machine (SVM), Random Forest (RF), and Gradient Boosting Decision Tree (GBDT).

[0018] The technical solution provided by this invention brings at least the following beneficial effects:

[0019] This invention combines multimodal information from a patient's MRI structural and functional images, fusing features such as morphology, low-frequency amplitude signals, voxel symmetry homotopy connectivity signals, functional network attributes, and functional network edges. This fully utilizes existing information to improve the accuracy of the auxiliary diagnostic model. Furthermore, to address the overfitting issue caused by limited datasets and a large number of features in medical images, this invention proposes using PCA for feature dimensionality reduction and filtering for feature selection. This reduces feature redundancy, and both methods effectively select significantly different features for classification and diagnosis, further improving the final classification accuracy. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the structure of an auxiliary diagnostic device for bipolar disorder provided in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0023] This invention utilizes machine learning methods for the auxiliary diagnosis of bipolar disorder. This addresses the current technical problems of low diagnostic rates and significant reliance on physician subjectivity in bipolar disorder diagnosis. The invention acquires structural magnetic resonance imaging (MRI) images and resting-state functional magnetic resonance imaging (fMRI) images of patients, employing statistical and graph theory analysis methods to identify brain regions showing significant differences between patients and normal controls. These brain regions are then used to train a machine learning classifier, resulting in a high-performing classifier for auxiliary diagnosis.

[0024] like Figure 1 As shown, the auxiliary diagnostic device for bipolar disorder provided in this embodiment of the invention includes: a structural image preprocessing unit, a functional image preprocessing unit, a structural image feature extraction unit, a functional image feature extraction unit, a feature fusion unit, a classifier training unit, and an auxiliary diagnostic output unit.

[0025] The structural image preprocessing unit is used to preprocess the input magnetic resonance structural image, including: segmenting the image according to different tissue structures, and outputting the segmented gray matter image to the structural image feature extraction unit.

[0026] The functional image preprocessing unit is used to preprocess the input MRI functional images, including: removing a specified number of time points at the start of the MRI functional image scan, arranging the corresponding scan layers in chronological order based on the machine's scanning sequence, performing correction processing using a preset head motion correction method, establishing the correspondence between the spatial positions of functional images and structural images (to facilitate functional image spatial standardization), and performing spatial standardization processing (to increase comparability, transforming all subjects' image data from the original space to a unified standard space), spatial smoothing processing, delinear drift processing, and filtering (to preserve the BOLD (brain oxygenation level dependent) signal); and inputting the preprocessed MRI functional image sequence into the functional image feature extraction unit;

[0027] The structural image feature extraction unit resizes the input gray matter image to match the resolution of a pre-set first brain region partitioning template (e.g., Anatomical Automatic Labeling, AAL atlas, which includes 116 regions, 90 in the cerebrum and 26 in the cerebellum). It then extracts gray matter from the image according to the brain region partitioning template, obtaining image blocks for each region. The unit measures the gray matter volume of each region and calculates the Pearson correlation coefficient between any two gray matter volumes. Based on a pre-set threshold, it performs binary classification on the Pearson correlation coefficient between two regions, assigning a value of 1 to strong Pearson correlations and a value of 1 to weak Pearson correlations. The Pearson correlation coefficient is set to 0. Based on the Pearson correlation coefficients of all brain regions, the corresponding adjacency matrix is ​​constructed. Each brain region is treated as a node. A weighted undirected brain structure network is constructed based on this adjacency matrix. The feature information of the brain structure network is extracted based on the specified network indicators (including node attributes (node ​​degree, degree centrality, betweenness centrality, etc.) and global network attribute information (isomatch, small-world property, feature path length, global efficiency, clustering coefficient, etc.)). The feature information is then formed into structural image features and input into the feature fusion unit after dimensionality reduction.

[0028] The functional image feature extraction unit includes a brain functional network feature extraction unit, a low-frequency amplitude signal (ALFF) feature extraction unit, and a voxel symmetric homotopy connectivity (VMHC) feature extraction unit.

[0029] The brain functional network feature extraction unit is used to resize each image in the input MRI functional image sequence. The resized image is aligned with the resolution of a pre-set second brain region partitioning template (which can be the same as or different from the first brain region partitioning template; the Dosenbacher map is preferred, as it contains 160 regions, providing more detailed brain segmentation than the AAL template, with 142 regions in the cerebrum and 18 in the cerebellum). The second brain region partitioning template is then used to partition each image in the MRI functional image sequence, obtaining the time series of each brain region. Partial correlation coefficients are then used to calculate the time series of any two brain regions. The correlation coefficients between time series data of brain regions are calculated, and the correlation coefficients are binarized using a specified density threshold. If the correlation coefficients are greater than or equal to the density threshold, they are binarized to 1; otherwise, they are binarized to 0. An adjacency matrix of brain regions is constructed based on the binarized correlation coefficients, with each brain region treated as a node. A weighted undirected brain functional network is then constructed based on this adjacency matrix. Feature information of the brain functional network is extracted based on specified network indicators (packet node attributes and global network attribute information). The extracted brain functional network features are then filtered to select features, and the selected brain functional network features are then input into the feature fusion unit.

[0030] The low-frequency amplitude signal feature extraction unit is used to extract the low-frequency amplitude signal features of the magnetic resonance functional image sequence, and then input the feature fusion unit after feature selection by filtering method.

[0031] The voxel symmetric homotopy connectivity feature extraction unit is used to extract voxel symmetric homotopy connectivity features from magnetic resonance functional imaging sequences, and the features are selected by filtering before being input into the feature fusion unit.

[0032] The feature fusion unit performs feature fusion (concatenation) on the input features to obtain the multi-modal features of the current object; if the current object is a training object, its multi-modal features are input to the classifier training unit; if the current object is an object to be assisted in diagnosis, its multi-modal features are input to the auxiliary diagnosis output unit.

[0033] The classifier training unit constructs a training dataset based on the multimodal features of the training objects: It sets a classification label (i.e., binary label) for bipolar disorder for each training object based on whether it belongs to the normal group or the disease group. Based on the multimodal features and classification labels of a certain number of training objects, it trains multiple different specified classifiers until a preset training termination condition is met. Then, it outputs the trained classifications to the auxiliary diagnostic output unit; each classifier is used to output the predicted binary classification result.

[0034] The auxiliary diagnosis output unit inputs the multi-modal features of the object to be diagnosed into each classifier of this unit, obtains the prediction results of each classifier based on its output, and uses a voting method to determine the diagnosis result of the current object to be diagnosed and outputs it.

[0035] Furthermore, when extracting brain functional network features, the brain functional network feature extraction unit extracts the topological attribute features of the brain functional network and extracts the specified network features of the network corresponding to the upper triangular matrix of the adjacency matrix of the brain region. However, when extracting network features of the network corresponding to the upper triangular matrix of the adjacency matrix of the brain region, the topological attribute features of the network are not included.

[0036] The beneficial effect of this invention is that by combining multimodal information from the patient's MRI structural and functional images, it fuses features such as the patient's morphology, ALFF signal, VMHC signal, functional network attributes, and functional network edges, making full use of existing information to improve the accuracy of the auxiliary diagnostic model. Simultaneously, to address the overfitting phenomenon caused by the limited dataset and large number of features in medical images, it proposes using PCA for feature dimensionality reduction and filtering methods for feature selection, thereby reducing feature redundancy. Using these two methods, it is possible to effectively select significantly different features for classification and diagnosis, resulting in a slight improvement in the final accuracy.

[0037] In this embodiment, training was first performed separately for each feature class to verify the effectiveness of these features. SVM, GBDT, and RF methods were selected for initial model training. The feature set in this embodiment mainly includes all previously used features, and the accuracy comparisons are shown in Table 1.

[0038] Table 1 Comparison of accuracy of various models trained based on different feature sets

[0039]

[0040] Because the number of subjects is small but the feature set has a high dimensionality, this embodiment employs two methods—feature extraction and feature selection—to optimize model training. Principal Component Analysis (PCA) is used for feature extraction, and filtering is used for feature selection.

[0041] PCA uses orthogonal transformations to linearly transform the observations of a series of potentially correlated variables, thus projecting them onto a series of linearly uncorrelated variables. The main steps of the algorithm are as follows:

[0042] (1) Assuming there are m n-dimensional data, arrange the original data into an n-row m-column matrix X;

[0043] (2) Take the mean of each row of data in X;

[0044] (3) Calculate the covariance matrix C. The calculation formula is shown below.

[0045]

[0046] (4) Find the eigenvalues ​​and eigenvectors of the covariance matrix C;

[0047] (5) Arrange the eigenvectors into a matrix from top to bottom according to the size of their corresponding eigenvalues, and take the first k rows to form matrix P;

[0048] (6) Y = PX is the data after dimensionality reduction to k dimensions.

[0049] Table 2 shows the changes in dimensionality features and the accuracy of each model before and after using PCA for dimensionality reduction.

[0050] Table 2 Comparison of accuracy of each model after PCA dimensionality reduction.

[0051]

[0052]

[0053] As shown in Table 2, after dimensionality reduction, the accuracy of each model either increased or remained unchanged. However, for all feature sets, although the feature dimension was reduced to 109, the overall accuracy decreased. This indicates that the original information may contain some information that is useless for classification, and this information was retained after compression, thus interfering with the results. The increase in gray matter volume indicates that the patient's morphology has indeed undergone some changes.

[0054] The filtering algorithm uses univariate statistical tests to achieve optimal feature selection. Each feature is tested to measure its relationship with the target variable y, and features with poor scores are removed. This embodiment uses mutual information to measure the relationship between two variables, retaining the top 100 features. In other words, this invention uses mutual information to measure the relationship between two variables, then retains a specified number of features with the highest scores as the feature selection results. The filtered features are then used for classification, and the accuracy results are shown in Table 3.

[0055] Table 3 Comparison of accuracy of each model after feature selection using mutual information.

[0056]

[0057] As shown in Table 3, mutual information has a good filtering function for the edges of the functional network. After filtering, using 100 features extracted from the edges of the functional network to train the model significantly increases the accuracy of all three models. After filtering the features in all feature sets, the highest increase in the classification accuracy of SVM is nearly 12 percentage points, reaching 82.1%. In this process, feature selection for gray matter volume causes the accuracy of the three models to remain the same or decrease. In summary, this embodiment uses PCA to reduce the dimensionality of the gray matter volume feature set and uses a mutual information-based filtering method to reduce the dimensionality of the union of other feature sets.

[0058] After performing dimensionality reduction on each feature set, this embodiment trains three models based on the obtained multimodal feature sets and integrates these three models as base learners using a voting method. When using the voting method, the weights of SVM, GBDT, and RF are set to 1, 1, and 1, respectively, based on the model performance above.

[0059] Table 4 describes the results of each model and the ensemble model after 10-fold cross-validation.

[0060] Table 4 Comparison of metrics for each model and the ensemble model's 10-fold cross-validation

[0061]

[0062] As shown in Table 4, after appropriate feature reduction and fusion, the accuracy of the three machine learning models all increased compared to the previous results. The model integrated using the voting method achieved an accuracy of 84.5% and a precision of 90%, effectively diagnosing patients with the disease. Therefore, based on the above conclusions, this embodiment ultimately chose to use the voting method to integrate the three models, thereby obtaining the final auxiliary diagnostic model.

[0063] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0064] The above descriptions are merely some embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. An auxiliary diagnostic device for bipolar disorder, characterized in that, include: The system includes a structural image preprocessing unit, a functional image preprocessing unit, a structural image feature extraction unit, a functional image feature extraction unit, a feature fusion unit, a classifier training unit, and an auxiliary diagnostic output unit. The structural image preprocessing unit is used to preprocess the input magnetic resonance structural image, including: segmenting the image according to different tissue structures, and outputting the segmented gray matter image to the structural image feature extraction unit; The functional image preprocessing unit is used to preprocess the input MRI functional images, including: removing a specified number of time points at the start of the MRI functional image scan, arranging the corresponding scan layers in chronological order based on the machine's scan sequence, performing correction processing using a preset head motion correction method, establishing the correspondence between the spatial positions of functional images and structural images, and performing spatial standardization, spatial smoothing, delinear drift removal, and signal filtering; and inputting the preprocessed MRI functional image sequence into the functional image feature extraction unit. The structural image feature extraction unit resizes the input gray matter image to match the resolution of a preset first brain region partitioning template. It then extracts gray matter from the image according to the template, resulting in image blocks for each brain region. The unit measures the gray matter volume of each region and calculates the Pearson correlation coefficient between any two regions. Based on a preset threshold, it performs binary classification on the Pearson correlation coefficient, assigning a value of 1 to strong correlations and 0 to weak correlations. An adjacency matrix is ​​constructed based on the assigned Pearson correlation coefficients for all brain regions. Each brain region is treated as a node, and a weighted undirected brain structure network is built based on this adjacency matrix. Finally, based on specified network metrics and global network attribute information, the unit extracts feature information from the brain structure network, forming structural image features. After dimensionality reduction, these features are input into the feature fusion unit. The functional image feature extraction unit includes a brain functional network feature extraction unit, a low-frequency amplitude signal feature extraction unit, and a voxel symmetric homotopy connectivity feature extraction unit; The brain functional network feature extraction unit is used to resize each image of the input MRI functional image sequence, ensuring the resized image size matches the resolution of a preset second brain region partitioning template. It then partitions each image of the MRI functional image sequence using the second brain region partitioning template, obtaining a time series for each brain region. The unit calculates the correlation coefficient between any two brain region time series using partial correlation coefficients, and binarizes the correlation coefficients using a specified density threshold. If the correlation coefficient is greater than or equal to the density threshold, it is binarized to 1; otherwise, it is binarized to 0. An adjacency matrix of brain regions is constructed based on the binarized correlation coefficients, treating each brain region as a node. A weighted undirected brain functional network is then constructed based on this adjacency matrix. Finally, the unit extracts feature information from the brain functional network based on specified network indicators, filters the extracted brain functional network features for feature selection, and inputs the selected brain functional network features into the feature fusion unit. The low-frequency amplitude signal feature extraction unit is used to extract the low-frequency amplitude signal features of the magnetic resonance functional image sequence, and then input the features into the feature fusion unit after feature selection by filtering method. The voxel symmetry homotopy connectivity feature extraction unit is used to extract voxel symmetry homotopy connectivity features from magnetic resonance functional imaging sequences, and the features are selected by filtering before being input into the feature fusion unit. The feature fusion unit fuses the input features to obtain the multi-modal features of the current object; if the current object is a training object, its multi-modal features are input to the classifier training unit; if the current object is an object to be assisted in diagnosis, its multi-modal features are input to the auxiliary diagnosis output unit. The classifier training unit constructs a training dataset based on the multimodal features of the training objects: It sets a classification label for bipolar disorder for each training object based on whether it belongs to the normal group or the disease group. Based on the multimodal features and classification labels of a certain number of training objects, it trains multiple different classifiers until a preset training termination condition is met. Then, it outputs the trained classifications to the auxiliary diagnostic output unit. Each classifier is used to output the predicted binary classification result. The auxiliary diagnosis output unit inputs the multi-modal features of the object to be diagnosed into each classifier of this unit, obtains the prediction results of each classifier based on its output, and uses a voting method to determine the diagnosis result of the current object to be diagnosed and outputs it.

2. The auxiliary diagnostic device for bipolar disorder as described in claim 1, characterized in that, When extracting brain functional network features, the brain functional network feature extraction unit extracts the topological attribute features of the brain functional network and extracts the specified network features of the network corresponding to the upper triangular matrix of the adjacency matrix of the brain region. However, when extracting network features of the network corresponding to the upper triangular matrix of the adjacency matrix of the brain region, the topological attribute features of the network are not included.

3. The auxiliary diagnostic device for bipolar disorder as described in claim 1 or 2, characterized in that, The structural image feature extraction unit uses principal component analysis to reduce the dimensionality of structural image features.

4. The auxiliary diagnostic device for bipolar disorder as described in claim 1 or 2, characterized in that, The classifiers used in the classifier training unit include three types: support vector machines, random forests, and gradient boosting decision trees.

Images