Brain disease classification system based on machine learning

A machine learning-based brain disease classification system using combined functional and structural MRI features effectively distinguishes MS from NMOSD, addressing misdiagnosis and improving treatment outcomes.

CN119068271AInactive Publication Date: 2024-12-03南昌大学第一附属医院

Patent Information

Application Number
CN202411573561.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-11-06
Publication Date
2024-12-03
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

Current machine learning models struggle to accurately distinguish between multiple sclerosis (MS) and neuromyelitis optic spectrum disorder (NMOSD) due to similar imaging features, leading to misdiagnosis and ineffective treatment.

Method used

A brain disease classification system using machine learning that combines functional and structural magnetic resonance imaging features, including gray matter volume, resting-state functional connectivity, and low-frequency amplitude features, to train a classification model that differentiates between MS and NMOSD.

Benefits of technology

The system achieves improved accuracy in distinguishing between MS and NMOSD, reducing misdiagnosis and enabling appropriate treatment, thereby minimizing disability and social impact.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119068271A_ABST
    Figure CN119068271A_ABST
Patent Text Reader

Abstract

The invention relates to the field of medical image analysis, and provides a brain disease classification system based on machine learning, the system is used for realizing a brain disease classification method based on machine learning, and the system comprises the following steps: carrying out brain magnetic resonance detection on a target body to obtain resting state functional magnetic resonance data and a high-resolution three-dimensional T1 weighted image of the target body; carrying out characterization processing on the resting state functional magnetic resonance data and the high-resolution three-dimensional T1 weighted image to obtain multi-level image features; and inputting the multi-level image features into a pre-trained target classification model to obtain a classification result output by the target classification model, the classification result being used for indicating that the target body is an MS patient or is used for indicating that the target body is an NMOSD patient. The brain disease classification system is high in accuracy, and multiple sclerosis diseases and neuromyelitis optica pedigree diseases can be effectively identified.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of medical image analysis, and more specifically, to a brain disease classification system based on machine learning. Background Art

[0002] Multiple sclerosis (MS) and neuromyelitis optic spectrum disorder (NMOSD) are autoimmune diseases characterized by inflammatory demyelination of the white matter of the central nervous system, with similar imaging features. As MS and NMOSD patients progress over time, in addition to the extremely high risk of relapse, their condition will gradually worsen, and they may eventually develop disability. The latter was once considered a clinical subtype of MS and was even called "Asian MS." But in fact, the two differ in immune pathogenesis, clinical manifestations, treatment, and prognosis. When NMOSD patients are misdiagnosed as MS, immunomodulatory drug treatment worsens the patient's symptoms. Therefore, in order to reduce the degree of disability, economic difficulties, and social impact of MS and NMOSD patients, accurate diagnosis and appropriate treatment are essential.

[0003] Therefore, the present application provides a brain disease classification system based on machine learning to solve one of the above technical problems. Summary of the invention

[0004] The purpose of this application is to provide a brain disease classification system based on machine learning, which can solve at least one of the technical problems mentioned above. The specific solution is as follows: According to a specific embodiment of the present application, in a first aspect, the present application provides a brain disease classification system based on machine learning, wherein the system is used to implement a brain disease classification method based on machine learning, comprising: Perform brain magnetic resonance detection on a target body to obtain resting-state functional magnetic resonance data and high-resolution three-dimensional T1-weighted images of the target body; perform feature processing on the resting-state functional magnetic resonance data and the high-resolution three-dimensional T1-weighted images to obtain multi-level image features; wherein the multi-level image features include one or a combination of gray matter volume features, resting-state functional connectivity features, low-frequency amplitude features, and local consistency features; input the multi-level image features into a pre-trained target classification model to obtain a classification result output by the target classification model, and the classification result is used to indicate that the target body is an MS patient, or is used to indicate that the target body is an NMOSD patient.

[0005] In one embodiment, the target classification model is trained in the following manner: performing brain magnetic resonance imaging on the MS patient group and the NMOSD patient group, respectively, to obtain resting-state functional magnetic resonance imaging data and high-resolution three-dimensional T1-weighted images of the MS patient group and the NMOSD patient group, and performing brain region segmentation and feature extraction based on a preset brain map to obtain first feature data corresponding to the MS patient group and second feature data corresponding to the NMOSD patient group; comparing and screening the first feature data and the second feature data according to a preset similarity threshold to screen out the first feature data and the second feature data whose similarity is greater than the preset similarity threshold. feature data; for the remaining first feature data and second feature data after the comparison and screening, a minimum absolute shrinkage selection operator regression method is used for cross-validation to screen out the first feature data and second feature data with a coefficient value of zero to obtain target feature data; based on a second training cohort composed of the target feature data, a plurality of pre-configured different initial classification models are trained respectively to obtain a plurality of different initial classification models that have been trained; based on a second test cohort composed of the target feature data, the plurality of different initial classification models that have been trained are tested respectively, and the initial classification model that outputs the optimal classification result is used as the target classification model.

[0006] In one embodiment, the initial classification model is a support vector machine model or a logistic regression model; the preset brain atlas is an automatic anatomical labeling atlas or a Harvard-Oxford atlas.

[0007] In one embodiment, before comparing and screening the first feature data and the second feature data, the first feature data and the second feature data are randomly allocated to a first training queue and a first test queue according to a preset configuration ratio; wherein, in the preset configuration ratio, the first training queue has a configuration ratio of 70%, and the first test queue has a configuration ratio of 30%; wherein, the second training queue is composed of target feature data originally belonging to the first training queue, and the second test queue is composed of target feature data originally belonging to the first test queue.

[0008] In one embodiment, the system is also used to: preprocess the resting-state functional magnetic resonance data and high-resolution three-dimensional T1 weighted images of the MS patient group and the NMOSD patient group before performing brain region segmentation and feature extraction based on a preset brain map; wherein the preprocessing satisfies the requirement of improving the signal-to-noise ratio of the resting-state functional magnetic resonance data and high-resolution three-dimensional T1 weighted images of the MS patient group and the NMOSD patient group.

[0009] In one embodiment, the resting-state functional magnetic resonance data and high-resolution three-dimensional T1-weighted images of the MS patient group and the NMOSD patient group are preprocessed, including: converting the resting-state functional magnetic resonance data and high-resolution three-dimensional T1-weighted images from DICOM format to NIFIT format; for the data after format conversion, removing the data of the first ten time points arranged along the time series, and performing time layer correction; for the data after time layer correction, excluding the subjects whose maximum head movement exceeds 2.0 mm displacement or exceeds 2 degrees in angular movement to complete head motion correction; performing delinear trend processing and linear regression on the data after head motion correction to obtain head motion parameters, cerebrospinal fluid signals, and white matter signals; performing individual registration between the resting-state functional magnetic resonance data and the high-resolution three-dimensional T1-weighted images, segmenting and registering the high-resolution three-dimensional T1-weighted images to the Montreal Neurological Institute standard space, and resampling the registered data to a voxel size of 3.0 mm×3.0 mm×3.0 mm; Hz bandwidth, bandpass filtering is performed on the resampled data to complete the preprocessing.

[0010] In one embodiment, the system is further used to: perform dimensionality reduction processing on the target feature data before respectively training a plurality of different pre-configured initial classification models.

[0011] In one embodiment, the system is further used to: determine the optimal classification result with reference to the target operating characteristic curve and the area under the curve of the test queue for the classification results output by each of the initial classification models.

[0012] In one embodiment, the multiple different initial classification models use a maximum likelihood function as a loss function, and use a gradient descent method to solve the optimal parameters of the model.

[0013] In one embodiment, the multi-level image features are composed of resting-state functional connectivity features, low-frequency amplitude features, and local consistency features; during the characterization process, the data is segmented into brain regions based on the Harvard-Oxford atlas; and the target classification model is a logistic regression model.

[0014] Compared with the prior art, the above scheme of the embodiment of the present application has at least the following beneficial effects: the brain disease classification system based on machine learning provided by the present application can realize the machine learning model based on functional magnetic resonance imaging features and / or structural magnetic resonance imaging features to distinguish between MS and NMOSD. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A flow chart of a brain disease classification method based on machine learning according to an embodiment of the present application is shown; Figure 2 A flow chart of a method for training a target classification model is shown; Figure 3 A flow chart of a method for preprocessing resting-state functional magnetic resonance data and high-resolution three-dimensional T1-weighted images of MS patient groups and NMOSD patient groups is shown. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.

[0017] The terms used in the embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit the present application. The singular forms "a", "said" and "the" used in the embodiments of the present application and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings, and "multiple" generally includes at least two.

[0018] It should be understood that the term "and / or" used in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.

[0019] It should be understood that, although the terms first, second, third, etc. may be used to describe in the embodiments of the present application, these descriptions should not be limited to these terms. These terms are only used to distinguish the descriptions. For example, without departing from the scope of the embodiments of the present application, the first may also be referred to as the second, and similarly, the second may also be referred to as the first.

[0020] As used herein, the words "if" and "if" may be interpreted as "at the time of" or "when" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrases "if it is determined" or "if (stated condition or event) is detected" may be interpreted as "when it is determined" or "in response to determining" or "when detecting (stated condition or event)" or "in response to detecting (stated condition or event)", depending on the context.

[0021] It should also be noted that the term "includes", "comprising" or any other variation thereof is intended to cover non-exclusive inclusion, so that a commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprising a ..." do not exclude the existence of other identical elements in the commodity or device including the elements.

[0022] It should be particularly noted that any symbols and / or numbers in the specification that are not marked in the accompanying drawings are not drawing marks.

[0023] Multiple sclerosis (MS) and neuromyelitis optic spectrum disorder (NMOSD) are autoimmune diseases characterized by inflammatory demyelination of the white matter of the central nervous system, with similar imaging features. As MS and NMOSD patients progress over time, in addition to the extremely high risk of relapse, their condition will gradually worsen, and they may eventually develop disability. The latter was once considered a clinical subtype of MS and was even called "Asian MS." But in fact, the two differ in immune pathogenesis, clinical manifestations, treatment, and prognosis. When NMOSD patients are misdiagnosed as MS, immunomodulatory drug treatment worsens the patient's symptoms. Therefore, in order to reduce the degree of disability, economic difficulties, and social impact of MS and NMOSD patients, accurate diagnosis and appropriate treatment are essential.

[0024] Currently, machine learning models mostly use structural magnetic resonance imaging features to distinguish between MS and NMOSD. Functional imaging features have the potential to distinguish between MS and NMOSD, but traditional statistical schemes of functional imaging features cannot distinguish between MS and NMOSD patients well.

[0025] In view of this, this application considers using functional imaging features and structural imaging features combined with a machine learning model to distinguish between MS and NMOSD.

[0026] The optional embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0027] The embodiment provided in this application is a brain disease classification system based on machine learning, which is used to implement any of the following embodiments of a brain disease classification method based on machine learning.

[0028] Combine the following Figure 1 The embodiments of the present application are described in detail.

[0029] Figure 1A flowchart of a brain disease classification method based on machine learning is shown, Figure 1 As shown, the process includes the following steps S101 to S103.

[0030] Step S101 , performing brain magnetic resonance imaging on a target body to obtain resting-state functional magnetic resonance imaging data and a high-resolution three-dimensional T1-weighted image of the target body.

[0031] In the present application, the target body can be understood as an individual to be tested for a disease. The target body can be an MS patient, an NMOSD patient or a healthy individual, and the present application does not limit this status.

[0032] In this application, resting state functional magnetic resonance imaging data (rs-fMRI) and high-resolution 3D T1-weighted imaging are both reference data that are helpful for the identification of MS or NMOSD.

[0033] Step S102, performing feature processing on the resting-state functional magnetic resonance data and the high-resolution three-dimensional T1-weighted image to obtain multi-level image features.

[0034] Among them, the multi-level imaging features include one or a combination of gray matter volume (GMV) features, resting-state functional connectivity (RSFC) features, amplitude of low frequency fluctuation (ALFF) features, and regional homogeneity (ReHo).

[0035] In this application, considering that different types of features reflect the differences between MS and NMOSD in different dimensions, different types of combinations are generated by including different types of features respectively. Therefore, 15 combinations of multi-level image features are provided as examples, namely: (1) ReHo, ALFF, RSFC single functional image features and GMV single structural image features. (2) Multi-level functional image feature combination features: ReHo + ALFF, ReHo + RSFC, ALFF + RSFC, ReHo + ALFF + RSFC. (3) Functional image feature combined with structural feature combination: ReHo + GMV, ALFF + GMV, RSFC + GMV, ReHo + RSFC + GMV, ReHo + ALFF + RSFC + GMV, ReHo + ALFF + RSFC + GMV.

[0036] Among them, the combination of multi-level imaging features can improve the classification performance. RSFC reflects the temporal correlation between neurophysiological activities in different brain regions in space; ALFF reflects the intensity of spontaneous oscillation functional activities in the brain; ReHo, as a local brain connectivity measure, can detect the temporal consistency of neighboring regions of neural intrinsic activity; GMV reflects the changes in brain gray matter volume. Different types of features can provide complementary information to highlight the neurological differences between MS and NMOSD patients.

[0037] Step S103, inputting the multi-level image features into the pre-trained target classification model to obtain the classification result output by the target classification model.

[0038] In the present application, the classification result is used to indicate that the target is an MS patient, or is used to indicate that the target is an NMOSD patient.

[0039] The brain disease classification system based on machine learning provided in the present application can realize the distinction between MS and NMOSD based on a machine learning model based on functional magnetic resonance imaging features and / or structural magnetic resonance imaging features.

[0040] In the present application, fMRI can non-invasively detect the human brain in a waking state and showing behavioral activities. By tracking the signals of the whole brain in a variety of cognitive and behavioral states, or mapping the differences associated with specific characteristics or clinical symptoms, fMRI has enhanced our understanding of brain function and its association with normal and abnormal behavior. And rs-fMRI refers to a functional magnetic resonance imaging scan performed on a target body, and these targets are instructed to let their minds wander (i.e., think about anything that appears in their minds) and stare at the crosshairs (or close their eyes). No other motor, perceptual or cognitive processes are explicitly required. Since no specific tasks need to be performed, confusion caused by cognitive or executive dysfunction in patients with MS and NMOSD is avoided.

[0041] For example, the present application can train the target classification model in the following manner: For the convenience of description and distinction, the feature data corresponding to the MS patient group is referred to as the first feature data, and the feature data corresponding to the NMOSD patient group is referred to as the second feature data.

[0042] Figure 2 A flow chart of a method for training a target classification model is shown, such as Figure 2 As shown, the process includes the following steps S201 to S205.

[0043] Step S201, performing brain magnetic resonance imaging on the MS patient group and the NMOSD patient group, respectively, to obtain resting-state functional magnetic resonance imaging data and high-resolution three-dimensional T1-weighted images of the MS patient group and the NMOSD patient group, and performing brain region segmentation and feature extraction based on a preset brain map to obtain first feature data corresponding to the MS patient group and second feature data corresponding to the NMOSD patient group.

[0044] Step S202: comparing and screening the first feature data and the second feature data according to a preset similarity threshold, so as to screen out the first feature data and the second feature data whose similarity is greater than the preset similarity threshold.

[0045] Step S203, cross-validating the remaining first feature data and second feature data after comparison and screening by using the least absolute shrinkage selection operator regression method to screen out the first feature data and second feature data with zero coefficient values ​​to obtain target feature data.

[0046] In this application, cross-validation is performed by the least absolute shrinkage selection operator regression (LASSO) method to obtain the meaningful features after screening.

[0047] Step S204: Based on the second training queue formed by the target feature data, the pre-configured multiple different initial classification models are trained respectively to obtain multiple different initial classification models that have been trained.

[0048] Step S205, based on the second test queue composed of the target feature data, the multiple different initial classification models that have been trained are tested respectively, and the initial classification model that outputs the best classification result is used as the target classification model.

[0049] In the present application, the first feature data and the second feature data obtained after brain region segmentation and feature extraction based on the preset brain map respectively characterize the multi-level image features of MS patients and NMOSD patients, and respectively include one or a combination of ReHo, ALFF, RSFC, and GMV. Accordingly, the characterization processing involved in the above step S102 of the present application is similar or identical to steps S201 to S203 in the present embodiment, and correspondingly includes steps such as brain region segmentation and feature extraction. The specific processing method can refer to the embodiment of the present application or any of the following embodiments, which will not be repeated here.

[0050] In this application, through feature screening, the features most relevant to the target variable can be identified and retained, while those that contribute little or no to the model performance can be eliminated. Correlation analysis, as a method for quantifying the relationship between features, can help us understand the underlying patterns and interactions in the data. By analyzing the correlation between features, features with high predictive power and independence can be selected more accurately, thereby improving the effect and results of feature selection. The least absolute shrinkage and selection operator (LASSO) regression compresses the model by adding an L1 regularization term (i.e., the sum of the absolute values ​​of the variable coefficients) to the loss function, reducing some coefficients to zero, thereby achieving variable selection.

[0051] As a feasible embodiment, the present application can implement screening based on a preset similarity threshold in the following manner.

[0052] For example, each feature of the MS and NMOSD patient groups was first z-score normalized. Further, the Pearson product-moment (Pearson) correlation test was used to evaluate the linear correlation between features to weaken multicollinearity and eliminate redundant features. If two features have a strong correlation, they will have a higher absolute value of the correlation coefficient. When the Pearson correlation coefficient between a feature was found to be >0.65, one of the features was selected for subsequent analysis. Based on the AAL template, after feature screening based on the Pearson correlation coefficient threshold >0.65, a total of 3361 features (49%) were retained. These included 73 ALFF features (63%), 3211 RSFC features (48%), 12 GMV features (10%), and 65 ReHo features (56%). Based on the HOA template, after the same feature screening with a Pearson correlation coefficient threshold of >0.65, 3104 (47%) features were retained, including 60 (54%) ALFF features, 2951 (47%) RSFC features, 13 (12%) GMV features, and 80 (71%) ReHo features.

[0053] By way of example, the processing schemes for obtaining various types of feature data are shown below.

[0054] GMV: During the preprocessing process, the T1 image was segmented to obtain the GMV, and the GMV was spatially smoothed (Gaussian kernel: 6 mm) for further analysis and statistics.

[0055] ALFF: In preprocessing, the time series of each voxel is filtered (0.01Hz ~ 0.1 Hz) to remove the influence of low-frequency drift and high-frequency noise, such as breathing and heart rate. Then, the filtered time series is transformed into the frequency domain by fast Fourier transform to obtain the power spectrum. Since the power of a given frequency is proportional to the square of the amplitude of the frequency component of the original time series in the time domain, the square root is calculated at each frequency of the power spectrum, and the average square root of 0.01Hz ~0.1 Hz is obtained at each voxel to calculate the ALFF value. Then, the ALFF value is converted into a z distribution. Finally, the image obtained after z transformation is spatially smoothed (Gaussian kernel: 6mm) for further analysis and statistics; ReHo: In preprocessing, the time series of each voxel was filtered (0.01Hz ~ 0.1 Hz) to remove the influence of low-frequency drift and high-frequency noise, such as breathing and heart rate. Then, the Kendall concordance coefficient was used to calculate the similarity between a single voxel and the surrounding 26 voxels to obtain the KCC value of each voxel to form ReHo. Finally, the image obtained after z-transformation was spatially smoothed (Gaussian kernel: 6mm) for further analysis and statistics; RSFC: The average time series of each ROI was obtained by averaging the rs-fMRI time series of all voxels within the ROI. Subsequently, Pearson correlation analysis was performed on all pairs of ROI average time series, and the correlation coefficients were then converted to z values ​​using Fisher z transformation. Finally, the images obtained after z transformation were spatially smoothed (Gaussian kernel: 6 mm) for further analysis and statistics. A total of 6670 z-transformed correlation coefficients were obtained as RSFC between all ROI pairs.

[0056] As an optional embodiment, any initial classification model in the present application can be a support vector machine model (SVM) or a logistic regression model (LR), and the preset brain map referenced when performing brain region segmentation can be an automated anatomical labeling (AAL) or Harvard-Oxford Atlas (HOA).

[0057] In this application, machine learning is a method of completing specific tasks by allowing computers to automatically learn rules and patterns from data. Support vector machine is a classic supervised learning algorithm used to solve binary and multi-classification problems. The core idea is to classify by finding an optimal hyperplane in the feature space with the largest interval. SVM can perform linear or nonlinear classification, regression, and even outlier detection tasks. It is one of the most popular models in the field of machine learning. SVM is particularly suitable for the classification of small and medium-sized complex data sets. Logistic regression is a linear model widely used in binary classification problems. The linear regression results are mapped to the [0,1] interval through the Sigmod function to achieve the transformation of continuous variables to probabilities. The model uses the maximum likelihood function as the loss function and uses the gradient descent method to solve the optimal parameters. Compared with linear regression, logistic regression uses thresholds for classification rather than continuous prediction. Different brain segmentation schemes may produce regions of interest of different sizes and locations, resulting in unique feature representations that affect classification performance. The main purpose of the automatic anatomical labeling atlas is to provide a standard anatomical reference for fMRI data. However, the boundaries of each ROI in the atlas are depicted using sulcal landmarks, which are less consistent with the boundaries of cell structures. The Harvard-Oxford atlas contains information on the geometry of sulci and gyri, which may better reflect individual differences.

[0058] In the present application, before comparing and screening the first feature data and the second feature data, the first feature data and the second feature data are randomly allocated to the first training queue and the first test queue according to a preset configuration ratio. Among them, in the preset configuration ratio, the first training queue has a configuration ratio of 70%, and the first test queue has a configuration ratio of 30%. Among them, the second training queue is composed of the target feature data originally belonging to the first training queue, and the second test queue is composed of the target feature data originally belonging to the first test queue.

[0059] As a feasible implementation method, the MS patient group and the NMOSD patient group can be randomly allocated in advance to a first training cohort and a first test cohort in a ratio of 7:3, and then by performing feature screening in the first training cohort and the first test cohort, a first training cohort and a first test cohort containing only target feature data are obtained to serve as the second training cohort and the second test cohort.

[0060] As another feasible implementation, the MS patient group and the NMOSD patient group can be randomly allocated to the first training cohort and the first test cohort in a ratio of 7:3 in advance, and then the training labels are set for the first feature data and the second feature data in the first training cohort, and the test labels are set for the first feature data and the second feature data in the first test cohort. On this basis, the first feature data and the second feature data with corresponding labels are screened as a whole, and the training cohort and the test cohort are reset according to the corresponding labels for the target feature data obtained after the screening, so as to obtain the second training cohort and the second test cohort.

[0061] As an optional embodiment, the present application may also preprocess the resting-state functional magnetic resonance data and high-resolution three-dimensional T1-weighted images of the MS patient group and the NMOSD patient group before performing brain region segmentation and feature extraction based on the preset brain map. On this basis, the present application may perform brain region segmentation and feature extraction based on the preset brain map for the preprocessed resting-state functional magnetic resonance data and high-resolution three-dimensional T1-weighted images.

[0062] The preprocessing of the resting-state functional magnetic resonance data and the high-resolution three-dimensional T1-weighted images satisfies the requirement of improving the signal-to-noise ratio of the resting-state functional magnetic resonance data and the high-resolution three-dimensional T1-weighted images of the MS patient group and the NMOSD patient group.

[0063] As a feasible embodiment, the present application can complete the preprocessing of resting-state functional magnetic resonance data and high-resolution three-dimensional T1-weighted images in the following manner.

[0064] Figure 3A flowchart of a method for preprocessing resting-state functional magnetic resonance data and high-resolution three-dimensional T1-weighted images of MS patient groups and NMOSD patient groups is shown, as shown in FIG. Figure 3 As shown, the following steps are included.

[0065] The resting-state functional magnetic resonance imaging data and high-resolution three-dimensional T1-weighted images were converted from DICOM format to NIFIT format.

[0066] Step S301 , for the format-converted data, remove the data of the first ten time points arranged along the time sequence, and perform time layer correction.

[0067] Step S302: for the time-layer corrected data, exclude subjects whose maximum head movement exceeds 2.0 mm displacement or whose angular movement exceeds 2 degrees to complete the head movement correction.

[0068] Step S303, performing delinear trend processing and linear regression on the data after head motion correction to obtain head motion parameters, cerebrospinal fluid signals, and white matter signals.

[0069] Step S304, individual registration is performed between the resting-state functional magnetic resonance data and the high-resolution three-dimensional T1-weighted image, the high-resolution three-dimensional T1-weighted image is segmented and registered to the Montreal Neurological Institute standard space, and the registered data is resampled to a voxel size of 3.0 mm×3.0 mm×3.0 mm.

[0070] Step S305 , performing bandpass filtering on the resampled data within a bandwidth of 0.01 Hz to 0.1 Hz to complete preprocessing.

[0071] In this application, by performing a series of preprocessing steps on resting-state functional magnetic resonance data and high-resolution three-dimensional T1-weighted images, including format conversion, removal of the first ten time points, time layer correction, head motion correction, delinear trend processing, individual registration, spatial standardization and bandpass filtering, this application significantly improves the signal-to-noise ratio of the data, reduces motion artifacts and interference signals, and enhances the consistency and comparability of cross-individual data, thereby providing a high-quality data foundation for subsequent feature extraction and machine learning model construction, and ultimately improving the accuracy and reliability of the differential diagnosis of MS and NMOSD.

[0072] As a feasible embodiment, the present application can also perform dimensionality reduction processing on the target feature data before training the pre-configured multiple different initial classification models respectively. For example, the high-dimensional RSFC feature is further reduced in dimension according to the number of connections between the brain and the brain region, and the brain region with multiple RSFCs is selected, and the RSFC between them is retained as a feature to distinguish between MS and NMOSD patients. Based on the AAL template, 9 ALFF, 14 ReHo and 31 RSFC functional imaging features and 5 GMV structural imaging features are finally retained. Based on the HOA template, 8 ALFF features, 14 ReHo features, 4 GMV features and 31 RSFC features are finally retained to distinguish between MS and NMOSD patients. The high-dimensional RSFC feature is further reduced in dimension according to the number of connections between the brain and the brain region, and the brain region with multiple RSFCs is selected, and the RSFC between them is retained as a feature to distinguish between MS and NMOSD patients. Based on the AAL and HOA brain templates, 7 RSFCs are finally retained.

[0073] As a feasible implementation method, for the classification results output by each initial classification model, the target operating characteristic curve and the area under the curve of the test queue can be referred to to determine the optimal classification result.

[0074] For example, all initial classification models are verified in the internal test cohort, and the discriminative ability of the model is evaluated by the target operating characteristic (ROC) curve and AUC of the test cohort. Quantitative indicators include accuracy, sensitivity and specificity. The calculation formulas for accuracy, sensitivity and specificity are as follows: Accuracy = (TP + TN) / (TP + TN + FP + FN) * 100%; Sensitivity = TP / (TP+FN)*100%; Specificity=TN / (TN+FP)*100%.

[0075] True positive (TP): a positive sample that is identified as positive; true negative (TN): a negative sample that is identified as negative; false positive (FP): a negative sample that is mistakenly identified as positive; false negative (FN): a positive sample that is mistakenly identified as negative.

[0076] As an optional embodiment, in the present application, multiple different initial classification models use the maximum likelihood function as the loss function, and use the gradient descent method to solve the optimal parameters of the model.

[0077] In this application, based on the AAL brain atlas, 6902 multi-level resting-state functional imaging features and 116 GMV structural imaging features were extracted from MRI images of 56 cases of MS and 36 cases of NMOSD. After feature screening, 9 ALFF, 14 ReHo and 7 RSFC functional imaging features and 5 GMV structural imaging features were retained. The area under the curve (AUC) of the SVM and LR classifiers constructed by single MRI features ranged from 0.635 to 0.778 and 0.738 to 0.786, respectively. The SVM and LR models constructed by the combination of RSFC, ALFF and ReHo multi-level functional imaging features performed best, with AUCs of 0.857, 95% confidence interval (CI): 0.618-1.000; 0.929, 95% CI: 0.762-1.000, respectively, and the classification performance was better than that of single functional feature and structural feature models. The AUC ranges of the SVM and LR classifiers constructed by coupling functional-structural features were approximately: 0.738-0.825, 0.722-0.889, respectively, and the classification performance was lower than that of the model constructed by multi-level functional features (RSFC, ALFF, and ReHo). Based on the HOA brain atlas, the most discriminative features were extracted and screened to construct the classifier, indicating that the model constructed by the multi-level functional imaging features of RSFC, ALFF, and ReHo was the most robust model for distinguishing MS and NMOSD patients, and the functional imaging features combined with GMV structural imaging features did not significantly improve the classification performance (AUC of the SVM model: 0.875 vs. 0.759; AUC of the LR model: 0.917 vs. 0.870).

[0078] It can be seen that in the embodiment of the present application, based on data verification, when the multi-level image features are composed only of resting-state functional connectivity features, low-frequency amplitude features, and local consistency features, and the data is segmented into brain regions based on the Harvard-Oxford atlas, and a logistic regression model is used as the target classification model, the model has the best performance and the highest classification accuracy. Therefore, this embodiment is used as a preferred embodiment to achieve the best accuracy in the classification results of brain diseases of the target body.

[0079] On this basis, the brain disease classification system based on machine learning provided in this application can realize the machine learning model based on functional magnetic resonance imaging features to distinguish between MS and NMOSD, and has a better classification accuracy compared to the classification scheme of the combined structural magnetic resonance imaging features mentioned above in this application.

[0080] In addition, the present application can also perform statistical analysis on the clinical information and the meaningful features after screening of the MS patient group and the NMOSD patient group according to a preset test algorithm.

[0081] In a preferred embodiment, the step of collecting clinical and neuropsychological assessments of all MS patients and NMOSD patients includes: (1) Chinese version of Paced Auditory Serial Addition Test (PASAT): a neuropsychological test to assess the patient's cognitive function, especially information processing speed and working memory. The lower the score, the higher the degree of cognitive impairment; (2) Expanded Disability Status Scale (EDSS): an assessment of the current cumulative disability status of the disease. The higher the score, the more severe the neurological deficit. An EDSS score of ≤2.5 indicates a low disability state, an EDSS score between 3 and 6 indicates a moderate disability state, and an EDSS score of ≥6.5 indicates a severe disability state; (3) Modified Fatigue Impact Scale (MFIS): mainly measures the impact of fatigue on the cognitive, physical and psychosocial functions of MS and NMOSD patients. The higher the score, the higher the fatigue level.

[0082] In a preferred embodiment, in the step of performing statistical analysis on the clinical information and the meaningful features after screening of the MS patient group and the NMOSD patient group according to a preset test algorithm, correlation analysis is performed between the discriminative brain regions and clinical manifestations of the MS and NMOSD patient groups, and partial correlation analysis is performed with age and gender as covariates (P × n<0.05 after Bonferroni correction).

[0083] In addition, the present application exemplarily illustrates some other feasible embodiments below.

[0084] In this application, all MS and NMOSD subjects were evaluated with the following scales after MRI scanning: (1) Chinese version of Paced Auditory Serial Addition Test (PASAT): a neuropsychological test that evaluates the patient's cognitive function, especially information processing speed and working memory. The lower the score, the higher the degree of cognitive impairment; (2) Expanded Disability Status Scale (EDSS): an assessment of the current cumulative disability status of the disease. The higher the score, the more severe the neurological deficit. An EDSS score of ≤2.5 indicates a low disability state, an EDSS score between 3 and 6 indicates a moderate disability state, and an EDSS score of ≥6.5 indicates a severe disability state; (3) Modified Fatigue Impact Scale (MFIS): mainly measures the impact of fatigue on the cognitive, physical and psychosocial functions of MS and NMOSD patients. The higher the score, the higher the fatigue level.

[0085] In one example, MRI data (rs-fMRI images and high-resolution three-dimensional T1-weighted images) were acquired from 56 patients with MS and 36 patients with NMOSD using an eight-channel phased head coil Trio 3.0 T MRI scanner (Siemens Medical Solutions, Erlangen, Germany). The rs-fMRI image acquisition parameters were as follows: repetition time (TR) / echo time (TE) = 2000 / 30 ms, matrix = 64 × 64, field of view (FOV) = 210 mm × 210 mm, slice thickness = 4.00 mm, 240 time points, and scan time of 8 minutes and 6 seconds; the high-resolution three-dimensional T1-weighted image acquisition parameters were as follows: TR / TE = 1900 / 2.26 ms, matrix = 240 × 256, FOV = 215 mm × 230 mm, slice thickness = 1.00 mm, 176 sagittal slices, and scan time of 3 minutes and 34 seconds.

[0086] In one example of this application, the MRI images were preprocessed as follows: (1) the first 10 time points were removed; (2) time slice correction was performed; (3) head motion correction was performed, and subjects whose maximum head motion exceeded 2.0 mm displacement in any direction or exceeded 2 degrees in any angular motion were excluded; (4) linear trend removal was performed to remove interference signals, and linear regression was performed to obtain Friston-24 head motion parameters, cerebrospinal fluid and white matter signals (global signal regression was not performed); (5) spatial normalization: first, individual registration was performed between the functional image and the high-resolution T1 image, and then the T1 image was segmented and registered to the Montreal Neurological Institute standard space, and then resampled to a voxel size of 3.0 mm × 3.0 mm × 3.0 mm; (6) bandpass filtering was performed at 0.01-0.1 Hz to remove the influence of low-frequency drift and high-frequency noise, such as breathing and heart rate. After segmentation, the GMV map was obtained, and the GMV image was spatially smoothed (Gaussian kernel: 6 mm) for further analysis and statistics.

[0087] In one example of the present application, multi-level resting-state functional imaging features are calculated: (1) ALFF: After preprocessing, the time series of each voxel is filtered (0.01 Hz ~ 0.1 Hz) to remove the influence of low-frequency drift and high-frequency noise, such as breathing and heart rate. Then, the filtered time series is transformed into the frequency domain by fast Fourier transform to obtain the power spectrum. Since the power of a given frequency is proportional to the square of the amplitude of the frequency component of the original time series in the time domain, the square root is calculated at each frequency of the power spectrum, and the average square root of 0.01 Hz ~ 0.1 Hz is obtained at each voxel to calculate the ALFF value. Then, the ALFF value is converted into a z distribution. Finally, the image obtained after z transformation is spatially smoothed (Gaussian kernel: 6 mm) for further analysis and statistics. (2) ReHo: After preprocessing, the time series of each voxel is filtered (0.01 Hz ~ 0.1 Hz) to remove the influence of low-frequency drift and high-frequency noise, such as breathing and heart rate. Then, the Kendall concordance coefficient was used to calculate the similarity between a single voxel and the surrounding 26 voxels to obtain the KCC value of each voxel and form the ReHo map. Finally, the z-transformed images were spatially smoothed (Gaussian kernel: 6mm) for further analysis and statistics. (3) RSFC: The average time series of each ROI was obtained by averaging the rs-fMRI time series of all voxels within the ROI. Subsequently, Pearson correlation analysis was performed on all ROI average time series pairs, and the correlation coefficients were converted into z values ​​using Fisher z transformation. Finally, the z-transformed images were spatially smoothed (Gaussian kernel: 6mm) for further analysis and statistics. A total of 6670 z-transformed correlation coefficients were used as RSFC between all ROI pairs.

[0088] In an example of this application, feature extraction: AAL brain atlas was used to segment and extract the average ReHo value, average ALFF value, and average GMV value of 116 ROIs. A total of 6670 RSFCs were extracted, plus 116 ReHo and 116 ALFF, a total of 6902 multi-level resting state functional imaging features, and 116 GMV structural imaging features.

[0089] In one example of the present application, feature screening: all patients were randomly assigned to training cohorts and test cohorts in a ratio of 7:3. Each feature of MS and NMOSD patients in the training group was z-score normalized. Then, the Pearson correlation test was used to evaluate the linear correlation between the features to weaken multicollinearity and eliminate redundant features. If there is a strong correlation between two features, they will have a higher absolute value of the correlation coefficient. When the Pearson correlation coefficient between a feature is found to be >0.65, one of the features is selected for subsequent analysis. Based on the AAL template, after feature screening based on the Pearson correlation coefficient threshold >0.65, a total of 3361 features (49%) were retained. These included 73 ALFF features (63%), 3211 RSFC features (48%), 12 GMV features (10%), and 65 ReHo features (56%). The LASSO regression method was used for feature screening, and the non-zero coefficient was used as a valuable predictor for each feature, and the weight coefficient of the final selected feature was recorded. The high-dimensional RSFC features were further reduced in dimension according to the number of connections between the brain and brain regions, and brain regions with multiple RSFCs were selected, and the RSFC between them was retained as a feature to distinguish between MS and NMOSD patients. Based on the AAL template, 9 ALFF, 14 ReHo, and 31 RSFC functional imaging features and 5 GMV structural imaging features were finally retained. The high-dimensional RSFC features were further reduced in dimension according to the number of connections between the brain and brain regions, and brain regions with multiple RSFCs were selected, and the RSFC between them was retained as a feature to distinguish between MS and NMOSD patients. Finally, 9 ALFF, 14 ReHo, and 7 RSFC functional imaging features and 5 GMV structural imaging features were retained, as shown in Table 1.

[0090] Table 1 In Table 1, the English abbreviations are as follows: AAL, automated anatomical labeling; RSFC, resting-state functional connectivity; ALFF, amplitude of low frequency; ReHo, regional coherence; GMV, gray matter volume.

[0091] In an example of this application, the machine learning model: considering that different types of features reflect the differences between MS and NMOSD in different dimensions, different types of combinations are generated by including different types of features. Then, Python3.7.12 and Scikit-learn were used to develop a machine learning classification model for each feature combination. The performance differences of the two machine learning classifiers, SVM and LR, under different feature combinations were compared. In addition, all models were validated in an internal test cohort. The discriminative ability of the model was evaluated by the target body operating characteristic (ROC) curve and AUC of the test cohort. Quantitative indicators include accuracy, sensitivity and specificity. Therefore, there are 15 combinations, namely: (1) single functional imaging features of ReHo, ALFF, RSFC and single structural imaging features of GMV. The AUC ranges of SVM and LR classifiers constructed by single MRI features are: 0.635~0.778, 0.738~0.786, respectively; (2) multi-level functional imaging feature combination features; ReHo + ALFF, ReHo + RSFC, ALFF + RSFC, ReHo + ALFF + RSFC. The SVM and LR models constructed by combining RSFC, ALFF and ReHo multi-level functional imaging features performed best, with AUCs of 0.857, 95% CI: 0.618-1.000; 0.929, 95% CI: 0.762-1.000, respectively. The classification performance was better than that of the single functional feature and structural feature models; (3) Functional imaging features combined with structural features: ReHo + GMV, ALFF + GMV, RSFC + GMV, ReHo + RSFC +GMV, ReHo + ALFF + GMV, ALFF + RSFC + GMV, ReHo + ALFF + RSFC + GMV. The AUC ranges of the constructed SVM and LR classifiers were approximately: 0.738~0.825, 0.722~0.889, respectively. The classification performance was lower than that of the model constructed by multi-level functional features (RSFC, ALFF and ReHo), as shown in Table 2.

[0092] Table 2 In Table 2, the English abbreviations are annotated as follows. MS, multiple sclerosis; NMOSD, neuromyelitis optica spectrum disorder; AAL, automated anatomical labeling; RSFC, resting-state functional connectivity; ALFF, low-frequency amplitude; ReHo, regional coherence; GMV, gray matter volume; AUC, area under the curve; ACC, accuracy; SN, sensitivity; SP, specificity; SVM, support vector machine; LR, logistic regression.

[0093] In one example of the present application, repeatability analysis: in addition to the AAL brain template, the HOA brain template is also used. According to the HOA brain template, the entire brain excluding the cerebellum can be divided into 112 brain regions, and the analysis method is basically the same as the AAL template. The results of applying different templates are compared from two aspects, one is the brain area selected under each template, and the other is the classification results. Based on the HOA template, after the same feature screening process mentioned above, 8 ALFF, 14 ReHo and 7 RSFC functional imaging features and 4 GMV structural imaging features are finally obtained, as shown in Table 3 below.

[0094] Table 3 In Table 3, the English abbreviations are as follows: RSFC, resting-state functional connectivity; ALFF, low-frequency amplitude; ReHo, regional coherence; GMV, gray matter volume.

[0095] In an example of this application, similar to the discriminative brain regions screened based on the AAL template, the most discriminative features extracted and screened were used to construct a classifier, indicating that the model constructed by the multi-level functional imaging features of RSFC, ALFF and ReHo was the most robust model for distinguishing MS and NMOSD patients, and the functional imaging features combined with GMV structural imaging features did not significantly improve the classification performance (AUC of SVM model: 0.875 vs. 0.759; AUC of LR model: 0.917 vs. 0.870), as shown in Table 4 below.

[0096] Table 4 In Table 4, the English abbreviations are annotated as follows. MS, multiple sclerosis; NMOSD, neuromyelitis optica spectrum disorder; HOA, automated anatomical labeling; RSFC, resting-state functional connectivity; ALFF, low-frequency amplitude; ReHo, regional coherence; GMV, gray matter volume; AUC, area under the curve, ACC, accuracy; SN, sensitivity; SP, specificity; SVM, support vector machine; LR, logistic regression.

[0097] In one example of the present application, the clinical significance of the discriminative features: correlation analysis was performed between the discriminative brain regions and clinical manifestations of MS and NMOSD patients, partial correlation analysis was performed with age and gender as covariates (P ×n<0.05 after Bonferroni correction), and the results were displayed using a heat map. Among the most discriminative features (RSFC, ALFF, and ReHo) for constructing the best classifier, the ALFF value of the right orbital inferior frontal gyrus was negatively correlated with the MFIS score of the MS patient group (r = -0.337, P × n<0.05 after Bonferroni correction); the ALFF value of the right cerebellum area 8 was negatively correlated with the MFIS score of the NMOSD patient group (r = -0.545, P × n<0.05 after Bonferroni correction); the RSFC value between the left orbital inferior frontal gyrus and the right cerebellum area 3 was negatively correlated with the PASAT score of the NMOSD patient group (r = -0.544, P × n<0.05 after Bonferroni correction); the RSFC value between the left caudate nucleus and the left superior temporal gyrus was negatively correlated with the EDSS score of the NMOSD patient group (r = -0.476, P × n<0.05 after Bonferroni correction).

[0098] In this application, the combination of multi-level functional imaging features can improve the classification performance. RSFC reflects the temporal correlation between neurophysiological activities in different brain regions in space; ALFF reflects the intensity of spontaneous oscillation functional activities in the brain; ReHo, as a local brain connectivity measure, can detect the temporal consistency of neighboring regions of neural intrinsic activity; GMV reflects the changes in gray matter volume. Different types of features can provide complementary information to highlight the neurological differences between MS and NMOSD patients.

[0099] Among them, the combination of multi-level fMRI indicators improved the classification performance compared with a single MRI indicator, indicating that RSFC, ALFF and ReHo indicators have complementary effects on the classification of MS and NMOSD patients. However, the classification model based on the combined indicators of fMRI and GMV performed poorly on the test set. This suggests that structural indicators (ie, GMV in this study) may interfere with the fMRI indicators for the classification of MS and NMOSD patients. In addition, in this application, in addition to the AAL brain template, the HOA brain template was also used to evaluate the consistency of the research results under different brain region templates. The results show that the model constructed by RSFC, ALFF and ReHo is the most robust model for distinguishing between MS and NMOSD patients.

[0100] In summary, this application successfully developed a more accurate and robust neural functional model (RSFC, ALFF, and ReHo), providing convincing evidence that the use of machine learning combined with unimodal, multi-level rs-fMRI data can highly accurately identify MS and NMOSD.

[0101] Currently, no published studies have suggested the use of multi-level functional imaging features to effectively identify patients with MS and NMOSD. This application introduces multi-level functional imaging biomarkers for the first time to identify MS and NMOSD, and explores the effectiveness of a single imaging feature model and a functional-structural fusion model. In the internal validation cohort, it was found that compared with the functional-structural fusion model, the machine learning model constructed with a single modality and multi-level functional features showed effectiveness in identifying patients with MS and NMOSD, improving the clinical application value of functional images.

[0102] Although operations are described in a particular order in the drawings, this should not be understood as requiring that the operations be performed in the particular order shown or in serial order, or that all shown operations be performed to achieve desired results. In certain circumstances, multitasking and parallel processing may be advantageous.

[0103] Any steps, operations or procedures described herein may be performed or implemented using one or more hardware or software modules, either alone or in combination with other devices. In one embodiment, the software module is implemented using a computer program product including a computer-readable medium containing computer program code, which can be executed by a computer processor to perform any or all of the described steps, operations or procedures.

[0104] The foregoing description of the implementation of the present application has been given for the purpose of illustration and description. The foregoing description is not exhaustive nor is it intended to limit the present application to the exact form disclosed, and various variations and modifications may exist according to the above teachings or may be obtained from the practice of the present application. These embodiments are selected and described in order to illustrate the principles of the present application and its practical application, so that those skilled in the art can utilize the present application in various embodiments and various modifications suitable for the specific purpose contemplated.

[0105] It can be further understood that, unless otherwise specified, “connection” includes a direct connection without other components between the two, and also includes an indirect connection with other components between the two.

[0106] It is further understood that, although the operations are described in a specific order in the drawings in the embodiments of the present application, it should not be understood as requiring the operations to be performed in the specific order or serial order shown, or requiring the execution of all the operations shown to obtain the desired results. In certain environments, multitasking and parallel processing may be advantageous.

[0107] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the invention disclosed herein. The present application is intended to encompass any modification, use or adaptation of the present application, which follows the general principles of the present application and includes common knowledge or customary techniques in the field of the present application that are not disclosed in the present application. The specification and embodiments are intended to be exemplary only, and the true scope and spirit of the present application are indicated by the following scope of rights.

[0108] It should be understood that the present application is not limited to the precise structures described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present application is limited only by the scope of the appended claims.

[0109] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A brain disease classification system based on machine learning, characterized in that: The system is used to implement a brain disease classification method based on machine learning, including: Performing brain magnetic resonance detection on the target body to obtain resting-state functional magnetic resonance data and high-resolution three-dimensional T1-weighted images of the target body; Performing feature processing on the resting-state functional magnetic resonance data and the high-resolution three-dimensional T1-weighted image to obtain multi-level image features; wherein the multi-level image features include one or a combination of gray matter volume features, resting-state functional connectivity features, low-frequency amplitude features, and local consistency features; The multi-level image features are input into a pre-trained target classification model to obtain a classification result output by the target classification model, wherein the classification result is used to indicate that the target body is an MS patient, or is used to indicate that the target body is an NMOSD patient.

2. The system according to claim 1, characterized in that The target classification model is trained in the following manner: Performing brain magnetic resonance imaging on the MS patient group and the NMOSD patient group, respectively, to obtain resting-state functional magnetic resonance imaging data and high-resolution three-dimensional T1-weighted images of the MS patient group and the NMOSD patient group, and performing brain region segmentation and feature extraction based on a preset brain atlas to obtain first feature data corresponding to the MS patient group and second feature data corresponding to the NMOSD patient group; According to a preset similarity threshold, the first feature data and the second feature data are compared and screened to screen out the first feature data and the second feature data whose similarity is greater than the preset similarity threshold; For the remaining first feature data and second feature data after comparison and screening, a minimum absolute shrinkage selection operator regression method is used to perform cross-validation to screen out the first feature data and second feature data with zero coefficient values ​​to obtain target feature data; Based on the second training queue formed by the target feature data, the pre-configured multiple different initial classification models are trained respectively to obtain multiple different initial classification models that have been trained; Based on the second test queue formed by the target feature data, the multiple different initial classification models that have been trained are tested respectively, and the initial classification model that outputs the optimal classification result is used as the target classification model.

3. The system according to claim 2, characterized in that The initial classification model is a support vector machine model or a logistic regression model; The preset brain atlas is an automatic anatomical labeling atlas or a Harvard-Oxford atlas.

4. The system according to claim 2, characterized in that Before comparing and screening the first feature data and the second feature data, the first feature data and the second feature data are randomly allocated to a first training queue and a first test queue according to a preset configuration ratio; Among them, in the preset configuration ratio, the first training queue has a configuration ratio of 70%, and the first test queue has a configuration ratio of 30%; The second training queue is composed of target feature data originally belonging to the first training queue, and the second test queue is composed of target feature data originally belonging to the first test queue.

5. The system according to claim 2, characterized in that The system is also used to: Before performing brain region segmentation and feature extraction based on the preset brain atlas, preprocessing the resting-state functional magnetic resonance data and high-resolution three-dimensional T1-weighted images of the MS patient group and the NMOSD patient group; The preprocessing satisfies the requirement of improving the signal-to-noise ratio of the resting-state functional magnetic resonance data and the high-resolution three-dimensional T1-weighted images of the MS patient group and the NMOSD patient group.

6. The system according to claim 5, characterized in that The preprocessing of the resting-state functional magnetic resonance data and high-resolution three-dimensional T1-weighted images of the MS patient group and the NMOSD patient group includes: Converting the resting-state functional magnetic resonance imaging data and high-resolution three-dimensional T1-weighted images from DICOM format to NIFIT format; For the data after format conversion, remove the data of the first ten time points arranged along the time series, and perform time layer correction; For the corrected data of the time layer, subjects with a maximum head movement exceeding 2.0 mm displacement or an angular movement exceeding 2 degrees were excluded to complete the head motion correction; Performing delinear trend processing and linear regression on the head motion corrected data to obtain head motion parameters, cerebrospinal fluid signals, and white matter signals; performing individual registration between the resting-state functional magnetic resonance data and the high-resolution three-dimensional T1-weighted images, segmenting and registering the high-resolution three-dimensional T1-weighted images to the Montreal Neurological Institute standard space, and resampling the registered data to a voxel size of 3.0 mm×3.0 mm×3.0 mm; The resampled data is band-pass filtered within a bandwidth of 0.01 Hz to 0.1 Hz to complete the preprocessing.

7. The system according to claim 2, characterized in that The system is also used to: Before respectively training the pre-configured multiple different initial classification models, the target feature data is subjected to dimensionality reduction processing.

8. The system according to claim 2, characterized in that The system is also used to: For the classification results output by each of the initial classification models, the optimal classification result is determined by referring to the target operating characteristic curve and the area under the curve of the test queue.

9. The system according to claim 2, characterized in that The multiple different initial classification models use the maximum likelihood function as the loss function, and use the gradient descent method to solve the optimal parameters of the model.

10. The system according to claim 1, characterized in that The multi-level image features are composed of resting-state functional connectivity features, low-frequency amplitude features, and local consistency features; During the characterization process, the data is segmented into brain regions based on the Harvard-Oxford atlas; The target classification model is a logistic regression model.

Citation Information

Patent Citations

  • Image convolutional neural network disease prediction system based on multi-modal magnetic resonance imaging

    CN115359045A

  • Method and system for classifying resting-state functional magnetic resonance image data and related products

    CN115457334A

  • Method and device for processing and analyzing resting-state functional magnetic resonance image

    CN117672467A

Cited By

  • Medical image-based spinal cord injury analysis method and system

    CN120510140A

  • A spinal cord injury analysis method and system based on medical images

    CN120510140B