A Classification Method for Mild Traumatic Brain Injury Based on Multi-modal Image Feature Fusion

Abstract

A classification method for mild traumatic brain injury based on multi-modal image feature fusion. Aiming at the problem that the current lack of an objective auxiliary diagnosis method for mTBI is caused by the incomplete single-modal image information and the unclear features highly related to mTBI, it fuses the cortical volume, thickness, and surface area features extracted from TI structural images and the FA, MD, AD, RD, OD, ICVF, and ISOVF features extracted based on bundles from dMRI. On the one hand, in order to obtain features highly related to mTBI, the features with significant differences between groups after FDR correction are selected as the initially screened features. On the other hand, in order to reduce the feature dimension and make the model achieve a better fitting effect, the initially obtained features are used as the input of the logistic regression model, and finally the optimal learning model trained with the features with non-zero weights is obtained. The generalization performance of the model is evaluated by the test set evaluation index. Through this model, mTBI disease prediction is carried out on visitors, so as to provide an objective auxiliary diagnosis method.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing, and particularly to a classification method for mild traumatic brain injury based on multi-modal image feature fusion. Background Art

[0002] Mild traumatic brain injury (mTBI) refers to a traumatic brain injury with a loss of consciousness of less than 30 minutes, and its complications include chronic headache, dizziness, inattention, irritability, and impulsivity. Even professional medical staff generally believe that these symptoms will disappear within three months after the trauma. In fact, 20% of mTBI patients will continue to have the above symptoms, which often leads to the patients not receiving timely nursing treatment after the injury. mTBI patients have a high risk of neurodegenerative diseases and are prone to chronic traumatic encephalopathy in the case of repeated injuries. mTBI can change brain function, thus affecting various fields such as long-term cognition, neuropsychiatry, and social function. Patients may even experience suicide, depression, and post-traumatic stress disorder symptoms.

[0003] Conventional diagnostic imaging methods such as computed tomography (CT) and magnetic resonance imaging (MRI) lack sensitivity to the subtle anatomical structure abnormalities of mTBI. Some neuroimaging techniques used in research fields can show the structural and functional changes related to mTBI. Vergara et al. obtained rsFNC-based features using group independent component analysis and the correlation between resting state networks, and then used a linear support vector machine for classification. The above method is vulnerable to the influence of patients' emotional cognition during fMRI acquisition, making it difficult to obtain objective data. Li et al. found based on sMRI that changes in cortical thickness and surface area can be detected in the brain, and these changes are related to the white matter macroscopic structure, microscopic structure integrity, functional network connectivity changes, and cerebral blood flow in mTBI patients. Some researchers found that DTI has good sensitivity to the group-level abnormalities in people with traumatic brain injury. The indexes derived from DTI can describe the diffusion characteristics of white matter fiber bundles and infer the direction and process information of white matter fiber bundles from this. So far, the diagnosis of mTBI mainly relies on the subjective self-report of patients' clinical symptoms and lacks an objective auxiliary diagnosis method. On the one hand, it is mainly because the imaging features highly related to this disease are not yet clear. On the other hand, whether it is the method based on dMRI or the research based on sMRI generally uses single-modal images and lacks some important information in other modalities. Summary of the Invention

[0004] To overcome the problem of the lack of an objective auxiliary diagnosis method for this disease caused by the incomplete single-modal imaging information and the unclear characteristics related to mTBI, the present invention proposes a method for classifying mild traumatic brain injury based on multi-modal imaging feature fusion. This method uses TI structural images and dMRI multi-modal data to make up for the deficiency of incomplete single-modal data information and also avoids the deficiencies in fMRI acquisition. In feature selection, features with significant differences between the mTBI and normal subject groups after being screened by logistic regression are used to train the model, making up for the problem of unclear features directly related to this disease. Using the trained model for disease prediction can provide an auxiliary diagnosis method for mTBI.

[0005] The technical solution adopted by the present invention to solve its technical problems is as follows:

[0006] A method for classifying mild traumatic brain injury based on multi-modal imaging feature fusion, the method comprising the following steps:

[0007] 1) Extract features based on T1 images: The image of each included subject contains T1 images and DTI images. The cortex of the T1 data is segmented using the latest 210-class atlas of brainetome as a template, and the cortex volume, thickness, and surface area are calculated and used as features.

[0008] 2) Extract features based on DTI images: The multi-shell multi-tissue constrained spherical deconvolution MSMT-CSD method and deterministic tracking are used to estimate the fiber direction and track the DTI data. The automatic segmentation of the fibers is performed by the WhitematterAnalysis toolbox. Finally, the fiber bundles are segmented into a set number of bilateral hemispheric fiber bundles and commissural bundles. The fractional anisotropy (FA), mean diffusivity (MD), axial diffusivity (AD), radial diffusivity (RD), and orientation dispersion index (ODI), nerve density index (NDI), and volume fraction of isotropic diffusion (FISO) parameter images are calculated, and the values of the parameter images are extracted based on anatomically significant fiber bundles as features. The features obtained in step 1) and step 2) are combined to construct an original data set.

[0009] 3) Data set preprocessing: The original data set is randomly divided into a training set and a test set by stratified sampling at a set ratio. The missing values of the training set are processed. A linear regression model is established from the data extracted from the normal subjects in the training set to exclude the influence of covariates on the data, and then the training set is normalized. The data preprocessing method for the test set is the same, but it needs to be based on the training set.

[0010] 4) Feature selection: Select the features with significant differences between the mTBI and normal subject groups screened by logistic regression in the training set, calculate the P-values between feature groups and perform FDR correction. Set a threshold to exclude the features without significant differences in the training set. The range of FDR correction is limited within the same type of parameters. For example, only perform FDR correction on all FA features once. After all types of parameters are selected, fit them with a logistic regression model, select the features with non-zero weights, and the test set retains the same features as the training set;

[0011] 5) Model training and evaluation: Use ten-fold cross-validation to put the training set obtained in step 4) into a machine learning model for training. After training, put the test set obtained in step 4) into the model for classification prediction, and compare it with the true label values to obtain evaluation indicators such as AUC to evaluate the generalization performance of the model, and select the best classifier according to the generalization performance.

[0012] The beneficial effects of the present invention are as follows: By using TI structural images and dMRI multi-modal data, it makes up for the defect of incomplete information in single-modal data and also avoids the deficiencies in fMRI acquisition. In feature selection, features with significant differences between the mTBI and normal subject groups screened by logistic regression are used to train the model, making up for the problem of unclear features directly related to the disease at present. Using the trained model for disease prediction can provide an auxiliary diagnosis method for mTBI. Brief Description of the Drawings

[0013] Figure 1 It is a schematic diagram of the steps of the present invention. Detailed Embodiments

[0014] In order to make the technical solution of the present invention clearer, the present invention will be further described below.

[0015] Refer to Figure 1 , a method for classifying mild traumatic brain injury based on multi-modal image feature fusion, the method comprising the following steps:

[0016] 1) Extract features based on T1 images: A total of 76 subjects were included in the present invention, including 42 patients and 34 normal subjects. The label "1" represents a patient, and the label "0" represents a normal subject. The images of each included subject contain T1 images and DTI images. The recon–all instruction of the freesurfer software performs complete segmentation of the cerebral cortex on the T1 image of each subject. BN_Atlas_freesurfer divides the cerebral cortex into 210 regions, and the segmentation template comes from Brainetome. Then, calculations are performed on the cerebral cortex of each brain region to obtain 210 values of cortical thickness, volume, and surface area as features. The calculated cortical thickness is the average value of the cortical thickness of the corresponding brain region;

[0017] 2) Extract features based on DTI images: The MSMT-CSD method is used to estimate fiber directions from DTI data and the deterministic tracking method is used to trace the fibers. The automatic segmentation of the fibers is performed by the WhitematterAnalysis toolkit. The steps include: 2.1) Register the DTI images to the same atlas through rigid and non-rigid transformations; 2.2) Divide the registered fiber bundles into 800 fiber bundles, including 716 bilateral hemisphere bundles and 84 commissural bundles; 2.) Apply the inverse transformation matrix including non-rigid and rigid ones to transform the fiber bundles back to the individual space. The NODDI parameters ODI, NDI, and FISO images based on multi-shell multi-b values are calculated by AMICO. The eigenvalues λ1, λ2, and λ3 in three directions can be obtained from the DTI tensor matrix. FA, MD, AD, and RD can be obtained through the following calculations:

[0018]

[0019]

[0020] AD = λ1

[0021]

[0022] Extract the values of the parameter images as features using the tract-based analysis method. Select the target fiber bundles with anatomical significance. The 100 centroids of the corresponding template bundles in the atlas are projected onto the registered target fiber bundles, and then each voxel in the target fiber bundle is assigned to the nearest point of the centroid. It can be divided into 100 segments in total. The average value of all voxels in the same segment of the target fiber bundle is used as a feature. Thus, 100 different parameter values can be extracted from each fiber bundle respectively. Fuse the features obtained in step 1) and step 2) to construct a dataset containing labels;

[0023] 3) Dataset preprocessing: Randomly and stratifiedly sample the dataset according to the labels. The number of samples in the obtained test set accounts for 25% (19 cases) of the number of samples in the dataset, and the rest are the training set (25 cases). Handle the missing values in the training set, exclude the features with a relatively large total number of missing features in the training set, and fill the features with a small number of missing values with "0". Establish a linear regression model for each feature in the training set, and exclude the influence of age, gender, and education level on the values. The linear regression model can be simply represented by the following formula:

[0024] V' = V - (aAge + bEdu + cSex + d)

[0025] where V' is the feature value after regression processing, V is the feature before regression processing, a, b, c, and d are the coefficients fitted by the regression model, and Age, Edu, and Sex are the age, education level, and gender of the subjects respectively.

[0026] Normalize each feature of the dataset, and the normalization formula is as follows:

[0027]

[0028] Among them, f' is the scaled feature value, f is the feature value before scaling, f max and f min are respectively the maximum and minimum values of this feature before scaling in the training set.

[0029] The test set retains the same features as the training set, fills in the missing values with "0", and the regression coefficients and the maximum and minimum values of the features before scaling are based on the training set;

[0030] 4) Feature selection: Calculate the P-value between feature groups and perform FDR correction. The range of FDR correction is limited to features of the same type. For example, only perform one FDR correction on the P-values generated by all FA features. The Benjamini and Hochberg method is used for correction. First, sort all P-values, and then calculate the corrected value through the following formula:

[0031] q = (p × m) ÷ k < α

[0032] Among them, q represents the corrected p-value, m is the number of tests, k is the rank of the p-value of this test among all test times, and α is the threshold, which takes the value of 0.05 in the present invention.

[0033] Exclude the features with no significant difference in the training set with a threshold of 0.05. After the above feature selection for all types of parameter features, fit them with a logistic regression model, and then select the features with non-zero output model weights again to obtain the input data of the machine learning model. The test set retains the same features as the training set;

[0034] 5) Model training and evaluation: Use ten-fold cross-validation to put the training set obtained in step 4) into the machine learning model for training, fine-tune the model parameters with the results of the validation set to obtain the optimal model. After training, put the test set obtained in step 4) into the model for classification prediction, and compare it with the true label value to obtain evaluation indicators such as AUC to evaluate the generalization ability of the model.

[0035] Figure 1 This is the schematic diagram of the steps of the present invention. As shown in the figure: Calculate the parameter map from the original DTI image, extract features based on the bundle method, obtain cortical information from the T1 structural image, fuse the DTI and cortical features, and then perform feature selection, training and evaluation of the model.

[0036] The content described in this specification is only an enumeration of the implementation forms of the inventive concept, and is only for illustrative purposes, and is not used to limit the patent scope of the present invention. All equivalent technical means that utilize the principles of the present invention and are conceived by those of ordinary skill in the art based on the inventive concept of the present invention should be included within the patent protection scope of the present invention.

Claims

1. A classification method for mild traumatic brain injury based on multi-modal image feature fusion, characterized in that The method includes the following steps: 1) Extract features based on T1 images: The images of each included subject contain T1 images and DTI images. The cortical segmentation of T1 data is performed using the brainetome 210-class atlas as a template, and the cortical volume, thickness, and surface area are calculated and obtained as features; 2) Extract features based on DTI images: The multi-shell multi-tissue constrained spherical deconvolution (MSMT-CSD) method and deterministic tracking are used to estimate the fiber orientation and track the DTI data. The automatic segmentation of fibers is performed by the WhitematterAnalysis toolkit. Finally, the fiber bundles are segmented into a set number of bilateral hemispheric fiber bundles and commissural bundles. The fractional anisotropy (FA), mean diffusivity (MD), axial diffusivity (AD), radial diffusivity (RD), orientation dispersion index (ODI), nerve density index (NDI), and volume fraction of isotropic diffusion (FISO) parameter images are calculated, and the values of the parameter images are extracted based on anatomically meaningful fiber bundles as features. The features obtained in steps 1) and 2) are combined to construct the original dataset; 3) Dataset preprocessing: The original dataset is randomly divided into a training set and a test set by stratified sampling at a set ratio. The missing values in the training set are processed. A linear regression model is established using the data extracted from the normal subjects in the training set to exclude the influence of covariates on the data, and then the training set is normalized. The data preprocessing method for the test set is the same, but it needs to be based on the training set; 4) Feature selection: Select the features with significant differences between the mTBI and normal subject groups in the training set after logistic regression screening. Calculate the P-values between the feature groups and perform FDR correction. Exclude the features without significant differences in the training set with a set threshold. The range of FDR correction is limited to the same type of parameters, and only one FDR correction is performed for all FA features. After all types of parameters are selected, they are fitted with a logistic regression model, and the features with non-zero weights are selected. The test set retains the same features as the training set; 5) Model training and evaluation: Use ten-fold cross-validation to put the training set obtained in step 4) into a machine learning model for training. After training, put the test set obtained in step 4) into the model for classification prediction, and compare it with the true label value to obtain evaluation indicators such as AUC to evaluate the generalization performance of the model, and select the best classifier according to the generalization performance; 2. A classification method for mild traumatic brain injury based on multi-modal image feature fusion according to claim 1, wherein The steps of step 2) include: 2.1) Register the DTI images to the same atlas through rigid and non-rigid transformations; 2.2) Divide the registered fiber bundles into 800 fiber bundles, including 716 bilateral hemispheric bundles and 84 commissural bundles; 2.3) Apply the inverse transformation matrix including non-rigid and rigid to transform the fiber bundles back to the individual space. Calculate the ODI, NDI, and FISO parameter images by AMICO. The eigenvalue λ1, λ2, and λ3 in three directions can be obtained from the DTI tensor matrix, and FA, MD, AD, and RD can be obtained through the following calculations: AD = λ1 The values of the parametric images are extracted as features using a beam-based analysis method. Target fiber bundles with anatomical significance are selected. The 100 centroids of the corresponding template bundles in the atlas are projected onto the registered target fiber bundles. Then, each voxel in the target fiber bundle is assigned to the nearest centroid point, and it can be divided into 100 segments in total. The average value of all voxels in the same segment of the target fiber bundle is used as a feature. Thus, 100 different parameter values can be extracted for each fiber bundle respectively. The features obtained in step 1) and step 2) are fused to construct a dataset containing labels.

3. A method for classifying mild traumatic brain injury based on multi-modal image feature fusion according to claim 1 or 2, characterized in that, In step 3), preprocess the dataset: perform random stratified sampling on the dataset according to the labels. The percentage of the number of test set samples in the number of dataset samples, and the rest is the training set. Process the missing values in the training set, exclude the features with a relatively large total number of missing features in the training set, and fill the features with a small number of missing values with "0". Establish a linear regression model for each feature in the training set, and exclude the influence of age, gender, and education level on the values. The linear regression model can be simply represented by the following formula: V' = V - (aAge + bEdu + cSex + d) where V' is the feature value after regression processing, V is the feature before regression processing, a, b, c, and d are the coefficients fitted by the regression model, and Age, Edu, and Sex are the age, education level, and gender of the subject respectively; Perform normalization processing on each feature of the dataset. The normalization formula is as follows: where f' is the scaled eigenvalue, f is the eigenvalue before scaling, f max and f min are the maximum and minimum values of this feature before scaling for the training set, respectively; The test set retains the same features as the training set, and the missing values are filled with "0". The regression coefficients and the maximum and minimum values of the features before scaling are based on the training set.

4. A method for classifying mild traumatic brain injury based on multi-modal image feature fusion according to claim 1 or 2, characterized in that, In step 4), calculate the P values between groups of features and perform FDR correction. The range of FDR correction is limited to the same type of features. Only perform FDR correction on the P values generated by all FA features once. The correction uses the Benjamini and Hochberg method. First, sort all P values, and then calculate the corrected value through the following formula: q = (p × m) ÷ k < α where q represents the corrected p value, m is the number of tests, k is the rank of the p value of this test among all test times, and α is the threshold; Exclude the features with no significant difference in the training set with a threshold of 0.

05. After the above feature selection for all types of parameter features, fit them with a logistic regression model, and select the features with non-zero output model weights again. Thus, the input data of the machine learning model is obtained. The test set retains the same features as the training set.

5. A method for classifying mild traumatic brain injury based on multi-modal image feature fusion according to claim 1 or 2, characterized in that In the above steps 1) and 2), using the latest template provided by Brainetome, the thickness, volume and surface area values of 210 cortices are calculated as features; in the DTI images, traditional DTI parameters FA, MD, AD, RD and more sensitive NODDI parameters ODI, NDI, FISO are extracted. On the 800 sub-classified fiber bundles, fiber bundles with anatomical significance are selected to extract features in a segmented form. Each feature is the average value of the parameter on the specified segment of the fiber bundle; the FA, MD, AD, RD and ODI, NDI, FISO features obtained from DTI are fused with the cortical thickness, volume and surface area features obtained from T1 to construct an initial dataset with multimodal information. This feature dataset contains the diffusion information of fibers and the gray matter information of the cerebral cortex.

6. A method for classifying mild traumatic brain injury based on multi-modal image feature fusion according to claim 1 or 2, characterized in that In the above step 4), considering both the clinical significance of mTBI and the effect of the machine learning model, two different methods of combining significant difference comparison and logistic regression are used for feature selection. First, calculate the P value between feature groups and perform FDR correction, and limit the range of FDR correction among features of the same type. Features with no significant difference in the training set are excluded with a threshold of 0.

05. This step is beneficial for screening out regions with abnormal brain structure or function in mTBI patients compared to those without mTBI; after the above feature selection, a logistic regression model is used for fitting, and features with non-zero output model weights are selected again to obtain the input data of the machine learning model. This step further reduces the number of features, obtains features linearly related to the disease, and prevents overfitting of the subsequent machine learning model.

Images