A method for assessing mild cognitive impairment based on key fiber bundles

Abstract

A method for evaluating mild cognitive impairment based on key fiber bundles. Since AD ​​patients are already in the middle and late stages when symptoms appear, and existing treatment methods are difficult to achieve effective results and can only delay the progression of the disease, the evaluation of mild cognitive impairment (MCI is in the intermediate stage between health and AD) is of great significance. The present invention selects diffusion tensor imaging magnetic resonance data as the main research object. This modality is a special form of magnetic resonance imaging and is currently the only non-invasive means to effectively observe and track brain white matter fiber bundles. It reflects multiple diffusion properties in brain white matter tissue. At the same time, unlike the previous pixel-level feature extraction of the entire magnetic resonance image, the present invention innovatively adopts fiber bundle-level features, that is, extracting key fiber bundles with significant differences between the mild cognitive impairment patient group and the healthy control group for feature fusion, which is a supplementary means to the traditional mild cognitive impairment evaluation method and can assist in the evaluation of mild cognitive impairment based on existing technologies.

Description

Technical Field

[0001] The present invention belongs to the field of medical image processing, and in particular relates to a method for assessing mild cognitive impairment based on key fiber bundles. Background Art

[0002] Alzheimer's disease (AD) is a neurodegenerative disorder characterized by progressive cognitive impairment and memory loss. It is currently the most common form of dementia. Because AD patients are already in the middle or late stages of the disease when symptoms appear, and existing treatments are rarely effective, only delaying the progression of the disease, the assessment of mild cognitive impairment (MCI), which is an intermediate stage between health and AD, is of great significance.

[0003] Magnetic resonance imaging (MRI) is a technology that non-invasively displays brain structure. Its rapid development has provided important support for understanding the physiological structure and functional connectivity of the brain. Many scholars and experts have analyzed and studied multimodal MRI data in an attempt to obtain more accurate MCI assessment results. In the article "Study on Imaging Prediction of Mild Cognitive Impairment Transformed into Alzheimer's Disease", the authors extracted brain gray matter from structural MRI images and then calculated the voxels of interest for classification prediction; in the article "Classification of Early Mild Cognitive Impairment Using rs-fMRI with Transfer Learning Feature Extraction", the authors extracted time series from resting-state functional MRI data and combined it with a transfer learning algorithm for classification prediction, achieving better accuracy than traditional algorithms.

[0004] Unlike the aforementioned classification and prediction methods, this present invention uses diffusion tensor imaging (DTI) magnetic resonance imaging (MRI) data as its primary research target. This modality is a specialized form of MRI and currently the only effective non-invasive method for observing and tracking white matter fiber tracts in the brain, reflecting multiple diffusion properties within the brain's white matter tissue. Furthermore, unlike previous approaches that extract pixel-level features from the entire MRI image, this present invention innovatively utilizes fiber tract-level features, extracting key fiber tracts that show significant differences between patients with mild cognitive impairment and healthy controls for feature fusion. This method complements traditional MCI assessment methods and can aid in the assessment of MCI using existing technologies. Summary of the Invention

[0005] To overcome the shortcomings of the existing technology, the present invention proposes a method for assessing mild cognitive impairment based on key fiber bundles. It selects diffusion tensor imaging magnetic resonance data as the main research object and innovatively adopts fiber bundle-level features, that is, extracts key fiber bundles with significant differences between the mild cognitive impairment patient group and the healthy control group for feature fusion. This is a supplementary means for traditional mild cognitive impairment assessment methods and can assist in the assessment of mild cognitive impairment based on existing technologies. Specifically, it includes:

[0006] Step (1) using the fiber automatic quantification method to track the whole-brain fiber bundles of the diffusion magnetic resonance images of the mild cognitive impairment patient group and the healthy control group, eliminating the free fiber bundles to obtain a preliminary whole-brain fiber bundle set, and then using the region of interest and probability map of the fiber bundle to eliminate the erroneous fibers and obtain the whole-brain white matter fiber bundle set;

[0007] Step (2) resampling each fiber in each key fiber bundle to 100 equidistant nodes to quantify the diffusion characteristics of the central part of the fiber bundle, calculating multiple characteristic indices, and then statistically analyzing the differences in fiber bundles between the two groups to obtain the key fiber bundle groups with significant differences;

[0008] Step (3) multiple characteristic indices of the fiber bundle groups with differences are expressed as feature vectors, and then all the extracted feature vectors are fused at the feature level using principal component analysis to obtain a multidimensional feature vector for each sample;

[0009] Step (4) divides the samples into a training set and a test set, selects a support vector machine (SVM) with a radial basis kernel function, and trains the SVM classifier using the feature vectors of the training set. After the training is completed, the feature vectors of the test set are input into the SVM classifier for classification evaluation, and the evaluation results are evaluated.

[0010] Furthermore, the step (1) specifically includes:

[0011] The tracking method used in the fiber automatic quantification method in step (1.1) is a deterministic tracking method: given one or more starting points, a new tracking direction is searched according to a set calculation method, and it is continuously propagated forward until certain termination conditions are reached, and finally a fiber streamline trajectory is obtained. For the termination conditions, that is, when the anisotropy index value of the whole-brain fiber bundle is less than 0.2 or the fiber bundle bending angle is greater than 45°, the fiber bundle is stopped from being tracked, and free fiber bundles that do not meet the tracking conditions are eliminated to obtain a preliminary whole-brain fiber bundle set;

[0012] Step (1.2) uses the region of interest and probability map of the fiber bundle to eliminate erroneous fibers and obtain the whole-brain white matter fiber bundle set. The specific steps are as follows:

[0013] Using the starting and ending areas of the fiber bundles as the regions of interest, the fibers that pass through the two regions of interest at the same time are screened out from the whole-brain fibers. Then, based on the probability maps of each key fiber bundle, the erroneous fibers are eliminated to obtain the whole-brain white matter fiber bundle set.

[0014] Furthermore, the step (2) specifically includes:

[0015] Step (2.1) uses multiple characteristic indices to measure brain plasticity, namely, partial anisotropy index, mean diffusivity, axial diffusivity, and radial diffusivity, and uses a tensor model to calculate the eigenvalues ​​of the tensor within the voxel;

[0016] Step (2.2) statistically analyzes the differences in fiber bundles between the two groups and identifies fiber bundles with significant differences. The specific steps are as follows: use independent sample T test and Pearson test to analyze multiple characteristic indicators of the two groups of fiber bundles, calculate the p-value and perform FDR correction, and consider p < 0.05 as a statistically significant difference, thereby obtaining the key fiber bundle groups with significant differences.

[0017] Furthermore, the step (3) specifically includes:

[0018] Step (3.1) extracts the average values ​​of multiple characteristic indices of the key fiber bundles of the sample and forms the characteristic vector of the sample, then performs standardization on the average values, and calculates the eigenvalues ​​and eigenvectors of the correlation coefficient matrix;

[0019] Step (3.2) uses principal component analysis to perform feature-level fusion on all the feature vectors extracted above. When the cumulative contribution rate of n principal components exceeds 90%, these n principal components are selected to complete the feature-level fusion. Finally, each sample is represented by n multi-dimensional feature vectors.

[0020] Furthermore, the step (4) specifically includes:

[0021] A feature vector classifier was constructed based on the SVM method to realize the classification evaluation of mild cognitive impairment. The support vector machine with radial basis kernel function was selected. Combined with the v-fold cross-validation method, three classifiers were constructed with m-1, m and m+1 folds for training. Then the feature vectors of the test set were input into the three SVM classifiers for classification. The classifier with the highest prediction accuracy was taken as the optimal classification. At the same time, the sensitivity and specificity were calculated as the evaluation criteria of the evaluation results.

[0022] The beneficial effects of the present invention are:

[0023] 1. The present invention innovatively adopts fiber bundle-level features, that is, extracting key fiber bundles with significant differences between the mild cognitive impairment patient group and the healthy control group for feature fusion, which can effectively evaluate early cognitive impairment based on the diffusion characteristic parameters within the fiber bundles.

[0024] 2. The white matter fiber bundle features used in the present invention are first screened for significant differences before classification, making the features more unique and representative. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is an implementation flow chart of the present invention.

[0026] Figure 2 Schematic diagram of the fiber bundles extracted by the tracking algorithm of the present invention.

[0027] Figure 3 Schematic diagram of fiber bundle characteristic indicators with significant differences according to the present invention. DETAILED DESCRIPTION

[0028] The following describes the specific embodiments of the present invention in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention. It should be noted that the following description is only used to explain the present invention and is not intended to limit the present invention.

[0029] The overall flow chart of the method for assessing mild cognitive impairment based on key fiber bundles proposed in the present invention is as follows: Figure 1 As shown, the details are as follows:

[0030] Step (1) Utilizing the fiber automatic quantification technology, the diffusion magnetic resonance images of the mild cognitive impairment patient group and the healthy control group were used to track the whole-brain fiber bundles, and some free fiber bundles were eliminated to obtain a preliminary whole-brain fiber bundle set. Then, using the region of interest (ROI) and probability map of the fiber bundles, the erroneous fibers were eliminated to obtain the whole-brain white matter fiber bundle set. The specific steps are as follows:

[0031] The tracking method used in the fiber automatic quantification technology in step (1.1) is a deterministic tracking method. Its principle is: given one or more starting points, a new tracking direction is searched according to the set calculation method, and it is continuously propagated forward until certain termination conditions are reached, and finally a fiber streamline trajectory is obtained. Some of the termination conditions here are that the FA (Fractional Anisotropy) value of the whole-brain fiber bundle is lower than 0.2 or the fiber bundle bending angle is greater than 45°, then the fiber bundle is stopped from being tracked, and the free fiber bundles that do not meet the tracking conditions are eliminated to obtain a preliminary whole-brain fiber bundle set, such as Figure 2 shown.

[0032] Step (1.2) uses the region of interest (ROI) and probability map of the fiber bundle to eliminate erroneous fibers and obtain the whole-brain white matter fiber bundle set. The specific steps are as follows:

[0033] Using the starting and ending regions of fiber tracts as regions of interest (ROIs), we screened out fibers from the whole brain that passed through both ROIs. We then removed erroneous fibers based on the probability maps of each key fiber tract. The probability map is the JHU white matter tract atlas, which was obtained by manually segmenting and registering brain data from 28 healthy individuals and needs to be converted to the individual brain space before use. This process was performed on the group-averaged DTI dataset in the MNI standard space, with the starting and ending ROIs of the fiber tracts defined.

[0034] Step (2) Each fiber in each key fiber bundle was resampled to 100 equidistant nodes to quantify the diffusion characteristics of the central part of the fiber bundle, and multiple characteristic indices were calculated. Then, the statistical software SPSS was used to analyze the differences in fiber bundles between the two groups to obtain the key fiber bundle groups with significant differences, as follows:

[0035] Step (2.1) uses four indicators to measure brain plasticity: fractional anisotropy (FA), mean diffusivity (MD), axial diffusivity (AD), and radial diffusivity (RD). The tensor model is used to calculate the eigenvalues ​​λ1, λ2, and λ3 of the tensor within the voxel. The calculation formulas and meanings of the four indicators are as follows:

[0036] The partial anisotropy index refers to the proportion of the anisotropic component of water molecules in the entire diffusion tensor, and its range of variation is 0 to 1. 0 represents unrestricted diffusion, such as the FA value of cerebrospinal fluid is close to 0; for very regular and directional tissues, the FA value is greater than 0, such as the FA value of brain white matter fibers is close to 1.

[0037]

[0038] To comprehensively evaluate the diffusion conditions of a tissue voxel or region, the mean diffusivity (MD) must eliminate the influence of anisotropic diffusion and be expressed as a constant parameter, meaning that its variation is independent of the direction of diffusion. MD reflects the overall molecular diffusion level (the size of the mean ellipsoid) and the overall diffusion resistance. MD only indicates the magnitude of diffusion and is independent of the direction of diffusion. A larger MD indicates a greater number of free water molecules in the tissue.

[0039]

[0040] Axial diffusivity represents the main direction of diffusion and is defined as:

[0041] AD=λ1 (3)

[0042] The radial spread represents the mean of the remaining two minor directions and is defined as:

[0043]

[0044] Step (2.2) Use the statistical software SPSS to analyze the differences in fiber bundles between the two groups and find the typical fiber bundles with significant differences. The specific steps are as follows: Use the independent sample T test and Pearson test to analyze the four index values ​​of the two groups of test subjects, calculate the p-value and perform FDR correction, and consider p<0.05 (p is the abbreviation of p-value, that is, p value, which is a very important indicator in statistics, related to the significance level, used to verify or overturn the original hypothesis in statistics, thereby reflecting the significance or non-significance of the results) as having statistical differences, thereby obtaining the key fiber bundle groups with significant differences. The fiber bundle characteristic indicators with significant differences are as follows: Figure 3 shown.

[0045] Step (3) represents multiple characteristic indices of the fiber bundle groups with differences as feature vectors, and then uses principal component analysis to perform feature-level fusion on all the extracted feature vectors to obtain a multidimensional feature vector for each sample, as follows:

[0046] In step (3.1), the average values ​​of multiple indicators (FA, MD, etc.) of the key fiber bundles of the sample are extracted and formed into the characteristic vector of the sample, which is then standardized and the eigenvalues ​​and eigenvectors of the correlation coefficient matrix are calculated.

[0047] In step (3.2), all the extracted eigenvectors are fused at the feature level using principal component analysis (PCA). PCA is a multivariate statistical analysis method that linearly transforms multiple variables to select a smaller number of important variables. PCA extracts the corresponding principal components and eigenvectors. When the cumulative contribution of n principal components exceeds 90%, these n principal components are selected to complete the feature-level fusion. Ultimately, each sample is represented by n multidimensional eigenvectors.

[0048] Step (4) divides the samples into a training set and a test set, selects a support vector machine (SVM) with a radial basis kernel function, and trains the SVM classifier using the feature vectors of the training set. After the training is completed, the feature vectors of the test set are input into the SVM classifier for classification evaluation, and the evaluation results are evaluated as follows:

[0049] Based on the SVM method, a feature vector classifier is constructed to realize the classification evaluation of mild cognitive impairment. The data of the present invention belongs to the case of linear inseparability, so the SVM method adopted is to transform the original data into a high-dimensional space so that it can be linearly separable, and then classify it using a linear separable method: a support vector machine with a radial basis kernel function is selected. In order to reduce the overfitting of the data, three classifiers are constructed with m-1, m and m+1 folds (m represents the specific number of v folds) for training, and then the feature vectors of the test set are respectively input into the three SVM classifiers for classification. The classifier with the highest prediction accuracy is taken as the optimal classification, and sensitivity and specificity are calculated as the evaluation criteria for the evaluation results. wherein, the definitions of accuracy, sensitivity and specificity are as follows:

[0050] Accuracy

[0051] Sensitivity

[0052] Specificity

[0053] Among them, TP: True Positive, indicating that the correct prediction is the positive class; TN: True Negative, indicating that the correct prediction is the negative class; FP: False Positive, indicating that the wrong prediction is the positive class; FN: False Negative, indicating that the wrong prediction is the negative class.

[0054] The present invention has been introduced in detail above, but the description of the specific implementation methods is only used to explain the method of the present invention and its core ideas, so that technicians in this technical field can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation methods. For ordinary technicians in this technical field, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.

Claims

1. A method for assessing mild cognitive impairment based on key fiber bundles, characterized by: The following steps are involved: Step (1) using the fiber automatic quantification method to track the whole-brain fiber bundles of the diffusion magnetic resonance images of the mild cognitive impairment patient group and the healthy control group, eliminating the free fiber bundles to obtain a preliminary whole-brain fiber bundle set, and then using the region of interest and probability map of the fiber bundle to eliminate the erroneous fibers and obtain the whole-brain white matter fiber bundle set; Step (2) resampling each fiber in each key fiber bundle to 100 equidistant nodes to quantify the diffusion characteristics of the central part of the fiber bundle, calculating multiple characteristic indices, and then statistically analyzing the differences in fiber bundles between the two groups to obtain the key fiber bundle groups with significant differences; Step (3) multiple characteristic indices of the fiber bundle groups with differences are expressed as feature vectors, and then all the extracted feature vectors are fused at the feature level using principal component analysis to obtain a multidimensional feature vector for each sample; Step (4) the samples are divided into a training set and a test set, a support vector machine (SVM) with a radial basis kernel function is selected, the feature vectors of the training set are used to train the SVM classifier, and after the training is completed, the feature vectors of the test set are input into the SVM classifier for classification evaluation, and the evaluation results are evaluated; The step (1) specifically includes: The tracking method used in the fiber automatic quantification method in step (1.1) is a deterministic tracking method: given one or more starting points, a new tracking direction is searched according to a set calculation method, and it is continuously propagated forward until certain termination conditions are reached, and finally a fiber streamline trajectory is obtained. For the termination conditions, that is, when the anisotropy index value of the whole-brain fiber bundle is less than 0.2 or the fiber bundle bending angle is greater than 45°, the fiber bundle is stopped from being tracked, and free fiber bundles that do not meet the tracking conditions are eliminated to obtain a preliminary whole-brain fiber bundle set; Step (1.2) uses the region of interest and probability map of the fiber bundle to eliminate erroneous fibers and obtain the whole-brain white matter fiber bundle set. The specific steps are as follows: The starting and ending areas of the fiber bundles were set as regions of interest, and the fibers that passed through both regions of interest were screened out from the whole-brain fibers. Then, based on the probability maps of each key fiber bundle, the wrong fibers were eliminated to obtain the whole-brain white matter fiber bundle set. The step (2) specifically includes: Step (2.1) uses multiple characteristic indices to measure brain plasticity, namely, partial anisotropy index, mean diffusivity, axial diffusivity, and radial diffusivity, and uses a tensor model to calculate the eigenvalues ​​of the tensor within the voxel; Step (2.2) statistically analyzes the differences in fiber bundles between the two groups and identifies the fiber bundles with significant differences. The specific steps are as follows: use independent sample T-test and Pearson test to analyze multiple characteristic indicators of the fiber bundles in the two groups, calculate the p-value and perform FDR correction, and consider p < 0.05 as a statistically significant difference. In this way, the key fiber bundle groups with significant differences are identified. The step (3) specifically includes: Step (3.1) extracts the average values ​​of multiple characteristic indices of the key fiber bundles of the sample and forms the characteristic vector of the sample, then performs standardization on the average values, and calculates the eigenvalues ​​and eigenvectors of the correlation coefficient matrix; Step (3.2) uses principal component analysis to perform feature-level fusion on all the feature vectors extracted above. When the cumulative contribution rate of n principal components exceeds 90%, these n principal components are selected to complete the feature-level fusion. Finally, each sample is represented by n multi-dimensional feature vectors.

2. The method for assessing mild cognitive impairment based on key fiber bundles according to claim 1, wherein: The step (4) specifically includes: A feature vector classifier was constructed based on the SVM method to realize the classification evaluation of mild cognitive impairment. The support vector machine with radial basis kernel function was selected. Combined with the v-fold cross-validation method, three classifiers were constructed with m-1, m and m+1 folds for training. Then the feature vectors of the test set were input into the three SVM classifiers for classification. The classifier with the highest prediction accuracy was taken as the optimal classification. At the same time, the sensitivity and specificity were calculated as the evaluation criteria of the evaluation results.

Images