Cognitive impairment trajectory structure magnetic resonance classification method based on multi-space scales

Abstract

The application discloses a cognitive impairment development trajectory structural magnetic resonance classification method based on multiple space scales, and comprises the following steps: selecting a first-level brain atlas, splitting and merging the first-level brain atlas into second and third-level atlases according to spatial structure relationships of sub-brain regions; pre-processing a structural magnetic resonance image; extracting seven structural magnetic resonance features of each sub-brain region in the three levels; using the three-level atlases, constructing a three-space-scale intracerebral layer connection network through multiple linear regression, combining the three-space-scale features, and then considering the relationship among the three levels to construct an interlayer connection network with a space scale of three; selecting a connection with a difference from the obtained connection matrix, recursively eliminating the features to obtain prediction features, taking single-space-scale features, three-space-scale fused features and interlayer features of multiple space scales as inputs, respectively, and using a classifier to evaluate the feature performance; and finally, the optimal classification features obtained have a further understanding of the widely recognized brain region connection.

Description

Technical Field

[0001] This invention belongs to the field of medical signal processing technology, specifically relating to a structural magnetic resonance imaging classification method for the developmental trajectory of cognitive impairment based on multiple spatial scales. Background Technology

[0002] Analyzing neuroimaging data of the brain to examine changes in its structure, function, and metabolism is a major research direction for the classification and prediction of mild cognitive impairment (MCI) and Alzheimer's disease (AD). Studies have shown that AD-related brain structural changes may appear earlier than in MCI, and these changes can be detected using structural magnetic resonance imaging (MRI).

[0003] In structural magnetic resonance imaging (SMR)-based brain analysis, researchers generally focus on SMR's performance in measuring changes in volume, thickness, and other characteristics in areas such as the hippocampus, entorhinal cortex, and amygdala. Some researchers also construct brain networks, dividing the brain into different regions of interest (ROIs), using these ROIs as nodes in the network, and the degree of connectivity between different ROIs as edges for connectivity analysis. While SMR features are advantageous in revealing the intrinsic structure and morphology of brain tissue, current analytical methods mostly use only a single map, considering features at a single spatial scale. Some studies have used multiple maps, but these simply merge features extracted from multiple maps without providing deeper physiological interpretation. Summary of the Invention

[0004] In order to solve the technical problems existing in the brain structure signal processing in the prior art, the purpose of this invention is to provide a multi-spatial-scale cognitive impairment development trajectory structural magnetic resonance classification method.

[0005] To achieve the above objectives, the present invention employs the following technical solution:

[0006] A structural magnetic resonance imaging (MRI) classification method for the developmental trajectory of cognitive impairment based on multiple spatial scales, characterized by the following steps:

[0007] Step 1: Selection and creation of multi-level maps;

[0008] Step two, after obtaining the spectrum, use FreeSurfer to preprocess the structural magnetic resonance images, specifically including:

[0009] Step 201: Sequence inspection and data format conversion

[0010] Check the format and quality of the magnetic resonance image sequence and convert the data into the format required by FreeSurfer;

[0011] Step 202: Correction of magnetic field inhomogeneity

[0012] Automatically estimate and correct magnetic field inhomogeneities to improve the accuracy of subsequent analyses;

[0013] Step 203: Head motion correction and initial registration

[0014] Based on the surface features of the skull and cerebral cortex in the images, the magnetic resonance images are roughly registered with standard space;

[0015] Step 205: White matter signal standardization to preserve more local structural information;

[0016] Step 206: Hard tissue segmentation

[0017] The brain tissue was segmented using an adaptive combinatorial model, including the brain tissue, hippocampus, amygdala, etc.

[0018] Step 207: Reconstruction of the Cerebral Cortex

[0019] Based on the segmentation results, the entire cerebral cortex is reconstructed using an adaptive surface reconstruction algorithm;

[0020] Step 208: Initial cortical surface registration and smoothing

[0021] The reconstructed cortical surface was initially registered with standard space and then smoothed using a spherical mapping algorithm.

[0022] Step 209: Based on the cortical surface registration results, the intensity values ​​and color information of the magnetic resonance image are mapped onto the cortical surface to display the local structure and features of the cortex;

[0023] Step 210: Cortical Marking and Partition Statistics

[0024] Calculate the structural indicators for each region based on the zoning results;

[0025] Step 3: Extract seven structural magnetic resonance features from each sub-brain region in the three levels;

[0026] Step 4: Use multiple linear regression to construct intra-layer and inter-layer connectivity networks in the brain.

[0027] Step 5: First, select the connections with statistical differences from the obtained multiple connection matrices, then perform feature recursive elimination to obtain predicted features, and use a classifier to evaluate the feature performance.

[0028] According to the present invention, in step one, the selection and creation of the multi-level atlas are completed in standard space using FreeSurfer, that is, the first-level brain atlas is selected, and then split and merged into second and third-level atlases according to the spatial structural relationships of brain subregions; the specific steps are as follows:

[0029] Step 101: The first-level map is the finest map. The BrainnetomeAtlas is used to construct the surface network of the brain, and 36 subcortical regions are removed from the map. The remaining 210 brain regions are used as the first-level map.

[0030] Step 102: The 210 brain regions in the first level are split into independent cortical label files, and then merged into 40 and 12 cortical label files respectively according to the spatial structure relationship. Finally, cortical annotation maps containing 40 and 12 brain regions are generated, which are the second and third level brain atlases.

[0031] Furthermore, in step three, the steps for extracting seven structural magnetic resonance features from each sub-brain region across the three levels are as follows:

[0032] Step 301: In step 210, various structural indices were obtained, and six commonly used cortical features were selected: cortical surface area, curvature index, folding index, mean curvature, cortical thickness, and cortical volume.

[0033] Step 302: Calculate the local cyclotron exponent using Freesurfer as the seventh feature;

[0034] Step 303: Using the three brain maps generated in Step 1, map them from the standard space to the individual space of each subject, and extract seven cortical features for each subject.

[0035] Specifically, in step four, the steps for constructing the intra-layer and inter-layer connectivity networks of the brain using multiple linear regression are as follows:

[0036] Step 401: Screen abnormal subjects. When extracting brain region features, images that cannot extract certain features will appear. Delete such images.

[0037] Step 402: Using the 210*7 features obtained from the subjects under the first-level map as input, the 7 features of each brain region as dependent variables, and the 7 features of the remaining 209 brain regions as independent variables, calculate the regression matrix.

[0038] Step 403: Since gender and age have a certain influence on brain structure, multiple linear regression is used to remove the influence of the two covariates.

[0039] Step 404: Using the second and third level brain maps respectively, repeat steps 402-403 to obtain the intralayer brain network connectivity matrices of three scales: 210*210, 40*40, and 12*12.

[0040] Step 405: Construct intra-layer features, using a method similar to steps 402-403, ultimately obtaining the intra-layer brain network connectivity matrix 210*40+210*12+40*12;

[0041] Preferably, in step five, the connections with statistical differences are first selected from the obtained multiple connection matrices, and then feature recursive elimination is performed to obtain predicted features. The steps for evaluating the feature performance using a classifier are as follows:

[0042] Step 501: Due to the large number of features obtained from constructing the brain network, feature elimination is first performed using a filtering method. For the two groups—those stable with mild cognitive impairment within three years and those transitioning to Alzheimer's disease—connections with statistically significant differences are selected.

[0043] Step 502: Use the wrapping method to filter features and use feature recursion elimination. Select features with the number of features 5, 10, 20, 50, 75, 100 and 200 respectively as input features for the classifier and select the SVM classifier.

[0044] Step 503: Select the three-level connection matrix obtained in step 404, repeat steps 501-502 to obtain three features with a spatial scale of 1, and feed them into the classifier to evaluate the classification performance of the structural feature matrix at the three levels; then fuse and filter the features at the three scales, and feed them into the classifier again to evaluate the performance; select the inter-layer features with a spatial scale of 3 obtained in step 405, repeat steps 501-502, and feed them into the classifier to evaluate the performance; finally, fuse the intra-layer features and inter-layer features, and evaluate the classification performance of the fused features.

[0045] Step 504: The feature classification results are as follows: for the three intralayer brain networks, the highest accuracies are 87%, 63%, and 60%, respectively, and the optimal number of features are 50, 50, and 100, respectively; the features obtained after feature filtering by fusing features from the three layers only have connections at the first layer, so the result is the same as the result of using only the first layer; for the intralayer brain networks, the highest accuracy is 93%, and the optimal number of features is 50; for the fused features, the highest accuracy is 85%.

[0046] Step 505: The 50 most accurate interlayer connections are mostly connections spanning the left and right hemispheres, involving the first and second levels as well as the first and third levels. They can explain deeper mechanisms of connections between several brain regions that are often identified as being associated with cognitive impairment in some fields: for example, the left cingulate gyrus and the right inferior parietal lobe, the left cingulate gyrus and the left insula, the left ventral occipital cortex and the right postcentral gyrus, and the self-connections existing in the cingulate gyrus.

[0047] The technological innovation of the structural magnetic resonance imaging classification method for the developmental trajectory of cognitive impairment based on multiple spatial scales in this invention lies in:

[0048] First, using the Brainnetome map, two more macroscopic maps were generated, and the connectivity matrix was calculated for each. Using data from these three different spatial scales as features, we explored at which scale the brain structure showed the most significant changes.

[0049] Secondly, the study integrates data from three different spatial scales as features to explore whether directly splicing features from different scales can complement each other.

[0050] Third, considering the connections between multiple spatial scales, the connections between layers are directly calculated, and finally the inter-layer connection features with a spatial scale of 3 are obtained.

[0051] Finally, the features within and between layers are fused to explore whether features at different levels play a complementary role.

[0052] The features extracted by this method classify cognitive impairment as stable or worsening over three years, with the best-performing features achieving an accuracy of 92%, providing an effective reference method for clinical auxiliary diagnosis. Furthermore, because the three spatial scales selected by this method have real spatial connections, the optimal classification features explain deeper mechanisms connecting several brain regions commonly identified as being associated with cognitive impairment: the left cingulate gyrus and right inferior parietal lobe, the left cingulate gyrus and left insula, the left ventral occipital cortex and right postcentral gyrus, and the self-connections of the cingulate gyrus. Attached Figure Description

[0053] Figure 1 The diagrams show the three levels of brain mapping used. In the diagram, (a) represents the first level of mapping (brain region 210), (b) represents the second level of mapping (brain region 40), and (c) represents the third level of mapping (brain region 12).

[0054] Figure 2 This is a schematic diagram of a structural magnetic resonance brain network calculated through regression.

[0055] Figure 3 This is a schematic diagram of hierarchical features.

[0056] Figure 4 This is a performance evaluation chart for classifying data from 133 participants in the stable cognitive impairment group and 163 participants in the cognitive impairment transition group using the method of this invention.

[0057] Figure 5 The 50 features with the highest accuracy based on interlayer connectivity are involved in three spatial levels; the outermost circle is the third level with the fewest brain regions, the middle circle is the second level, and the small circle next to the second level is the first brain region involved.

[0058] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. Detailed Implementation

[0059] This embodiment presents a structural magnetic resonance imaging (MRI) classification method for the developmental trajectory of cognitive impairment based on multiple spatial scales, including the following steps:

[0060] 1) Multi-level atlas selection and creation were completed using FreeSurfer in standard space. The first-level brain atlas was selected, and the brain regions in the atlas were split according to the spatial structural relationships of the brain sub-regions, merging them into second and third-level atlases. The specific steps are as follows:

[0061] Step 101: The first-level atlas is the finest atlas, selected from the BrainnetomeAtlas. Since the aim is to construct a surface brain network, 36 subcortical regions were removed from the atlas, using the remaining 210 brain regions as the first-level atlas. The results are referenced... Figure 1 Left image;

[0062] Step 102: The 210 brain regions in the first level are split into independent cortical label files. These are then merged into 40 and 12 cortical label files respectively, based on their spatial structure. This generates cortical annotation maps containing 40 and 12 brain regions, which are the second and third level brain atlases. The results are referenced... Figure 1 China Library Figure 1 The image on the right.

[0063] 2) After obtaining the spectrum, use FreeSurfer to preprocess the structural magnetic resonance images. The specific steps are as follows:

[0064] Step 201: Sequence inspection and data format conversion

[0065] Check the format and quality of the magnetic resonance image sequence and convert the data into the format required by FreeSurfer;

[0066] Step 202: Correction of magnetic field inhomogeneity

[0067] Automatically estimate and correct magnetic field inhomogeneities to improve the accuracy of subsequent analyses;

[0068] Step 203: Head motion correction and initial registration

[0069] Based on the surface features of the skull and cerebral cortex in the images, the magnetic resonance images are roughly registered with standard space;

[0070] Step 205: Signal standardization to preserve more local structural information;

[0071] Step 206: Hard tissue segmentation

[0072] The brain tissue was segmented using an adaptive combinatorial model, including the brain tissue, hippocampus, amygdala, etc.

[0073] Step 207: Reconstruction of the Cerebral Cortex

[0074] Based on the segmentation results, the entire cerebral cortex is reconstructed using an adaptive surface reconstruction algorithm;

[0075] Step 208: Initial cortical surface registration and smoothing

[0076] The reconstructed cortical surface was initially registered with standard space and then smoothed using a spherical mapping algorithm.

[0077] Step 209: Based on the cortical surface registration results, the intensity values ​​and color information of the magnetic resonance image are mapped onto the cortical surface to display the local structure and features of the cortex;

[0078] Step 210: Cortical Marking and Partition Statistics

[0079] The structural indicators for each region can be calculated based on the zoning results.

[0080] 3) Extract seven structural magnetic resonance features from each sub-brain region across the three levels. The specific steps are as follows:

[0081] Step 301: In step 210, various structural indices were obtained, and six commonly used cortical features were selected: cortical surface area, curvature index, folding index, mean curvature, cortical thickness, and cortical volume.

[0082] Step 302: Calculate the local cyclotron exponent using Freesurfer as the seventh feature;

[0083] Step 303: Using the three brain maps generated in Step 1, map them from the standard space to the individual space of each subject, and extract seven cortical features for each subject.

[0084] 4) Construct intra-layer and inter-layer connectivity networks in the brain using multiple linear regression. The specific steps are as follows:

[0085] Step 401: Screen abnormal subjects. When extracting brain region features, images that cannot extract certain features will appear. Delete such images.

[0086] Step 402: Refer to Figure 2 Using the 210*7 features obtained by the subjects under the first level map as input, the 7 features of each brain region are the dependent variables, and the 7 features of the remaining 209 brain regions are the independent variables, the regression matrix is ​​obtained.

[0087] Step 403: Since gender and age have a certain influence on brain structure, multiple linear regression is used to remove the influence of the two covariates.

[0088] Step 404: Using the second and third level brain maps respectively, repeat steps 402-403 to obtain the intralayer brain network connectivity matrices of three scales: 210*210, 40*40, and 12*12.

[0089] Step 405: Construct intra-layer features, referring to... Figure 3 The method is similar to steps 402-403, and the final intralayer brain network connection matrix is ​​210*40+210*12+40*12;

[0090] 5) Perform feature filtering on the obtained multiple connection matrices, and evaluate the feature performance using a classifier. The specific steps are as follows:

[0091] Step 501: Due to the large number of features obtained from constructing the brain network, feature elimination is first performed using a filtering method. For the two groups—those stable with mild cognitive impairment within three years and those transitioning to Alzheimer's disease—connections with statistically significant differences are selected.

[0092] Step 502: Use the wrapping method to filter features and use feature recursion elimination. Select features with the number of features 5, 10, 20, 50, 75, and 100 respectively as input features for the classifier and select the SVM classifier.

[0093] Step 503: Select the three-level connection matrix obtained in step 404, repeat steps 501-502 to obtain three features with a spatial scale of 1, and feed them into the classifier to evaluate the classification performance of the structural feature matrix at the three levels; then fuse and filter the features at the three scales, and feed them into the classifier again to evaluate the performance; select the inter-layer features with a spatial scale of 3 obtained in step 405, repeat steps 501-502, and feed them into the classifier to evaluate the performance; finally, fuse the intra-layer features and inter-layer features, and evaluate the classification performance of the fused features.

[0094] Step 504: The feature classification results are as follows: for the three intralayer brain networks, the highest accuracies are 87%, 63%, and 60%, respectively, and the optimal number of features are 50, 50, and 100, respectively; the features obtained after feature filtering by fusing features from all three layers only contain connections from the first layer, therefore the result is the same as using only the first layer; for the intralayer brain networks, the highest accuracy is 93%, and the optimal number of features is 50; for the fused features, the highest accuracy is 85%; refer to... Figure 4 Accuracy, precision, recall, and F1 score are used to evaluate classification performance.

[0095] Step 505: Reference the 50 most accurate interlayer connections Figure 5It can reveal more connections across the left and right hemispheres, involving the first and second levels as well as the first and third levels, and can explain deeper mechanisms of connections between several brain regions that are often identified as being associated with cognitive impairment in some fields: such as the left cingulate gyrus and the right inferior parietal lobe, the left cingulate gyrus and the left insula, the left ventral occipital cortex and the right postcentral gyrus, and the self-connections that exist in the cingulate gyrus.

[0096] The structural magnetic resonance imaging (SMRI) classification method for the developmental trajectory of cognitive impairment presented in this embodiment involves features at different spatial scales:

[0097] (1) Data from three different spatial scales are directly fused as features for classification. Using BN maps and two more macroscopic levels of maps, seven magnetic resonance features are extracted for each map, the connectivity matrix is ​​calculated, and finally features with a spatial scale of 3 are obtained;

[0098] (2) Considering the relationship between multiple spatial scales, the relationship between levels is directly calculated, and finally the characteristics of the three spatial scales are obtained.

[0099] (3) Taking into account both intra-layer and inter-layer features, the features generated in (1) and (2) are fused together;

[0100] Finally, feature selection was performed to find the classification features with the best classification performance. The results show that:

[0101] (1) From the perspective of a single spatial scale, the classification results of the three-level internal features selected by this method show that the brain structure has more significant changes at a finer spatial scale and has better classification performance.

[0102] (2) The classification performance obtained by directly fusing features from three spatial scales is the same as that obtained by using the finest feature alone, indicating that the direct fusion of features from different spatial scales does not play a complementary role and cannot enhance the classification performance.

[0103] (3) The use of inter-layer features, i.e. features with a spatial scale of 3, effectively enhanced the classification performance, indicating that the developmental trajectory of cognitive impairment shows greater differences in inter-layer connections, providing new ideas for the analysis of magnetic resonance data of other diseases.

[0104] (4) The fusion of inter-layer and intra-layer features reduced the accuracy of the classifier, indicating that the fusion of the two scales did not complement each other, but instead damaged the original performance.

Claims

1. A multi-space scale based cognitive impairment trajectory structural magnetic resonance classification method, characterized in that, Follow these steps: Step 1: Selection and creation of multi-level maps; Step two, after obtaining the spectrum, use FreeSurfer to complete the structural magnetic resonance imaging and perform preprocessing, specifically including: Step 201: Sequence inspection and data format conversion Check the format and quality of the magnetic resonance image sequence and convert the data into the format required by FreeSurfer; Step 202: Correction of magnetic field inhomogeneity Automatically estimate and correct magnetic field inhomogeneities to improve the accuracy of subsequent analyses; Step 203: Head motion correction and initial registration Based on the surface features of the skull and cerebral cortex in the images, the magnetic resonance images are roughly registered with standard space; Step 205: White matter signal standardization to preserve more local structural information; Step 206: Hard tissue segmentation Brain tissue was segmented using an adaptive combinatorial model, including brain tissue, hippocampus, and amygdala. Step 207: Reconstruction of the Cerebral Cortex Based on the segmentation results, the entire cerebral cortex is reconstructed using an adaptive surface reconstruction algorithm; Step 208: Initial cortical surface registration and smoothing The reconstructed cortical surface was initially registered with standard space and then smoothed using a spherical mapping algorithm. Step 209: Based on the cortical surface registration results, the intensity values ​​and color information of the magnetic resonance image are mapped onto the cortical surface to display the local structure and features of the cortex; Step 210: Cortical Marking and Partition Statistics Calculate the structural indicators for each region based on the zoning results; Step 3: Extract seven structural magnetic resonance features from each sub-brain region in the three levels; Step 4: Use multiple linear regression to construct intra-layer and inter-layer connectivity networks in the brain. Step 5: First, select the connections with statistical differences from the obtained multiple connection matrices, then perform feature recursive elimination to obtain predicted features, and use a classifier to evaluate the feature performance.

2. The method of claim 1, wherein, The selection and creation of the multi-level atlas described in step one is completed using FreeSurfer in standard space. Specifically, a first-level brain atlas is selected, and then split and merged into second and third-level atlases according to the spatial structural relationships of brain subregions. This includes: Step 101: The first-level map is the finest map. The brain network map is used to construct the surface network of the brain, and 36 subcortical regions are removed from the map. The remaining 210 brain regions are used as the first-level map. Step 102: The 210 brain regions in the first level are split into independent cortical label files, and then merged into 40 and 12 cortical label files respectively according to their spatial structure relationship. Finally, cortical annotation maps containing 40 and 12 brain regions are generated, which are the second-level and third-level brain atlases.

3. The method as described in claim 1, characterized in that, In step three, seven structural magnetic resonance features of each sub-brain region in the three levels are extracted, specifically: Step 301: In step 210, various structural indices were obtained, and six commonly used cortical features were selected: cortical surface area, curvature index, folding index, mean curvature, cortical thickness, and cortical volume. Step 302: Calculate the local cyclotron exponent using Freesurfer as the seventh feature; Step 303: Using the three brain maps generated in Step 1, map them from the standard space to the individual space of each subject, and extract seven cortical features for each subject.

4. The method as described in claim 1, characterized in that, Step four: Construct intra-layer and inter-layer connectivity networks in the brain using multiple linear regression. The specific steps are as follows: Step 401: Screen abnormal subjects. When extracting brain region features, images that cannot extract certain features will appear. Delete such images. Step 402: Using the 210 obtained by the subject under the first-level map The seven features are used as inputs, the seven features of each brain region are used as dependent variables, and the seven features of the remaining 209 brain regions are used as independent variables. The regression matrix is ​​then calculated. Step 403: Since gender and age have a certain influence on brain structure, multiple linear regression is used to remove the influence of the two covariates; Step 404: Using the second and third level brain maps respectively, repeat steps 402-403 to obtain the intra-layer brain network connectivity matrix at three scales. 210, 40 40, 12 12; Step 405: Construct inter-layer features, using a method similar to steps 402-403, ultimately obtaining the inter-layer brain network connectivity matrix 210. 40+210 12+40 12.

5. The method as described in claim 4, characterized in that, Step five includes the following steps: Step 501: Since the number of features obtained from constructing the brain network is too large, the first step is to use a filtering method to eliminate features. For the two groups, those that remain stable in mild cognitive impairment within three years and those that have transitioned to Alzheimer's disease, connections with statistical differences are selected. Step 502: Use the wrapping method to filter features and use feature recursion elimination. Select features with the number of features 5, 10, 20, 50, 75, and 100 respectively as input features for the classifier and select the SVM classifier. Step 503: Select the three-level connection matrix obtained in step 404, repeat steps 501-502 to obtain three features with a spatial scale of 1, and feed them into the classifier to evaluate the classification performance of the structural feature matrix at the three levels; then fuse and filter the features at the three scales, and feed them into the classifier again to evaluate the performance; select the inter-layer features with a spatial scale of 3 obtained in step 405, repeat steps 501-502, and feed them into the classifier to evaluate the performance; finally, fuse the intra-layer features and inter-layer features, and evaluate the classification performance of the fused features. Step 504: The feature classification results are as follows: for the three intralayer brain networks, the highest accuracies are 87%, 63%, and 60%, respectively, and the optimal number of features are 50, 50, and 100, respectively; the features obtained after feature filtering by fusing features from the three layers only have connections at the first layer, so the result is the same as the result of using only the first layer; for the intralayer brain networks, the highest accuracy is 93%, and the optimal number of features is 50; for the fused features, the highest accuracy is 85%, and the optimal number of features is 50. Step 505: The 50 most accurate interlayer connections are more cross-hemisphere connections, involving the first and second layers as well as the first and third layers, and can explain the deeper mechanisms of connections between several brain regions that are often identified as being associated with cognitive impairment in some fields.

Images