Methods and apparatus for constructing brain disease classification models based on multimodal microstructure topological features

Abstract

The method and device for constructing a brain disease classification model based on a multi-modal microstructure topological feature provided by the embodiments of the present application introduce cortex surface data into a structural structure space, and construct structural connection data and functional connection data at a millimeter-level microstructure scale, thereby reducing the calibration range of a brain region, enabling fine-grained extraction of lesion characteristics within the brain region, and making up for the problem that existing brain disease classification models based on MRI technology and human brain connection groups cannot extract abnormal brain characteristics at a microstructure scale and cannot accurately extract lesion regions and lesion characteristics. Meanwhile, the embodiments of the present application can set a downsampling ratio, select the resolution of the microstructure scale and the resolution of a macro-scale brain segmentation template, fuse multi-scale and multi-modal features, effectively establish a disease commonality and specificity damage evaluation strategy, and improve the classification accuracy of the model of the embodiments of the present application. Before constructing a brain network, the mapping relationship of different modal spaces is obtained through translation and rotation transformation, and thus the brain network in the individual space and the feature information of the individual space are constructed, which is conducive to the extraction of disease commonality and specificity characteristics. In addition, the embodiments of the present application select a classification and prediction model as a relatively mature and simple machine learning and deep learning model, thereby avoiding the problem that the dimensions of human brain structural and functional connection data at a microstructure scale are large, which is not conducive to feature extraction, feature fusion and model training in actual application.

Description

Technical Field

[0001] This application relates to the field of human brain neuroscience research technology, and in particular to a method and apparatus for constructing a brain disease classification model based on multimodal microstructure topological features. Background Technology

[0002] In recent years, magnetic resonance imaging (MRI) has been widely used in research on functional abnormalities and classification prediction of brain diseases due to its non-invasive nature and the ability to observe structural / functional abnormalities caused by brain developmental abnormalities. Currently, commonly used MRI techniques for brain disease research include structural magnetic resonance imaging (sMRI), diffusion MRI imaging (dMRI), functional magnetic resonance imaging (fMRI), and magnetic resonance spectroscopy (MRS). Among these, a widely used multimodal neuroimaging analysis method is the graph theory-based brain connectomics approach. This method provides a resource for mapping neuropsychiatric symptoms, characterizing brain connectivity maps at multiple spatial scales, and describing information interactions between regions. Specifically, it abstracts and represents the brain using network analysis methods, reducing the complexity of neuron- and gene-based analyses; it segments the brain based on brain structure or function and extracts features related to brain structure and function; it extracts features that meet specific requirements based on multimodal and multi-feature analysis, and establishes an effective disease classification model.

[0003] Currently, constructed human brain network models can be divided into structural networks (or analogical networks) and functional networks. The connections between nodes can be further divided into structural connectivity (SC) and functional connectivity (FC). The structural network of a human brain network model is typically represented by a structural connectivity matrix (SC matrix), measured using diffusion tensor / diffusion spectroscopy (DTI / DSI) technology, reflecting the physiological structure of the brain. The functional network of a human brain network model is typically represented by a functional connectivity matrix (FC matrix), measured using resting-state functional magnetic resonance imaging (fMRI) technology, reflecting the statistical connectivity relationships between different nodes in the brain network under functional states. Utilizing the network's topological properties, such as node degree, shortest path length, local efficiency, small-world characteristics, and modularity, researchers can measure the ability of a patient's brain or parts of the brain to resist external interference, its information transmission capabilities, and its ability to separate and integrate functions, providing effective features for establishing brain disease classification models.

[0004] Nevertheless, because the aforementioned methods focus on macroscopic-scale research and lack regional-level microstructural information and multi-level information integration frameworks, the classification of brain diseases largely relies on large brain regions and the connectivity between them, rather than delving into the fine-grained features of lesions within brain regions. This results in low accuracy of the extracted features used in model classification, especially for brain diseases with unclear etiologies and pathogenesis, and diseases with similar clinical manifestations, such as brain diseases, Parkinson's disease, Alzheimer's disease, and their subtypes. Fortunately, compared with other neuroimaging techniques, MRI technology has the advantage of high spatial resolution. This provides a new method to address the current challenge of low spatial resolution in brain research by constructing ultra-high resolution brain networks based on MRI, laying the foundation for improving the performance of brain disease classification models. Among existing techniques, methods for analyzing the brain based on ultra-high resolution structure employ voxel-based morphological analysis (VBM), specifically quantitatively calculating and analyzing changes in gray and white matter density or volume of each voxel in MRI to assess structural differences between healthy subjects and patients. Nevertheless, the features extracted by this method are limited to morphological features, and it cannot extract network topological features at a deep level or establish structural and functional connections in a fine-grained manner, which limits the performance of the final brain disease classification model.

[0005] Therefore, there is an urgent need to improve existing brain disease classification models based on brain connectomics to address the problem of inaccurate extraction of lesion regions and lesion features. Compared to the limited information observed by single-modal imaging techniques, multimodal imaging techniques, with their ability to compensate for the specific information between different modalities and improve the classification ability of the model, have become one of the key means to construct brain disease classification models. Summary of the Invention

[0006] This application provides a method and apparatus for constructing a brain disease classification model based on multimodal microstructure topological features. It aims to overcome the limitations of existing brain disease classification models based on MRI technology and the human brain connectome, which struggle to extract microstructural-scale abnormal brain features and exhibit inaccurate extraction of lesion regions and features. This model reveals structural and functional abnormalities in the diseased brain at the regional microstructural scale, enabling fine-grained feature mining of lesions within brain regions. By integrating multimodal brain structural and functional topological features at both macroscopic and microstructural levels and utilizing a variety of selectable machine learning prediction models, the accuracy of the brain disease classification model can be improved. This classification model can be applied to the early diagnosis and adjuvant treatment of brain diseases, particularly those with unclear etiologies and pathogenesis, and diseases with similar clinical manifestations, such as brain diseases, Parkinson's disease, Alzheimer's disease, and their subtypes.

[0007] Firstly, this application provides a method for constructing a brain disease classification model based on multimodal microstructure topological features, the method comprising:

[0008] Step 110: Obtain the brain disease patient {PAT} from a pre-set human brain neuroimaging database. n1 and healthy subjects {CON} n2 Multimodal imaging data, including T1-weighted images, dMRI data, and fMRI data; where {…} represents a set; n1 and n2 represent the number of patients with brain diseases and healthy subjects, respectively; a standardized rating scale {SCALE} for assessing symptoms of brain diseases was obtained. n1 、{SCALE} n2 ; Obtain the patient's duration of illness {DUR} n1 ; Obtain brain partition templates at the macroscopic scale of the standard spatial (MNI).

[0009] Step 120: Preprocess the acquired T1 data, dMRI data and fMRI data using neuroimaging processing software to obtain cortical surface data (pia mater.pial) and fiber tracking data (.trk).

[0010] Step 130: Use FreeSufer software to obtain the transformation matrix M from diffusion magnetic resonance space (dMRI) to structural space (T1).d2t The transformation matrix M from functional magnetic resonance imaging (fMRI) to structural T1. f2t The transformation matrix M from volume space to surface space T1 st1 ; Obtain the transformation matrix from the MNI space brain partition template to the individual structural space T1, and map the MNI space template to the individual T1 space; Use the individual space brain partition template to remove non-cortical tissue data from T1 cortical spatial fiber tracking data and fMRI data.

[0011] Step 140: Create an individual space template (Temp). ind The segmented regions of interest (ROIs) are mapped to the cortical space; the smallest unit of cortical pia mater data is obtained and used as the initial smallest unit (ROI) for microstructural segmentation. Δ And calculate the ROI Δ Area A Δ , establish roi Δ Mapping roi_id to its corresponding label Δ .

[0012] Step 150: Downsample and reconstruct the acquired cortical pia mater file, and use the new triangular mesh units as the final minimum unit (ROI) of the microstructure ROI. newΔ Calculate the ROI of this microstructure newΔ Area A newΔ And establish a mapping roi_id between the new microstructure and its corresponding label. newΔ .

[0013] Step 160: Based on T1 cortical spatial fiber tracking data trk st1 and fMRI data ts st1 Construct a binary structure connectivity matrix SC at the microstructure scale. b Weighted structure connection matrix SC w and functional connection matrix FC.

[0014] Step 170: Based on ROI newΔ Mapping roi_id to its corresponding label newΔ Structural connectivity matrix and functional connectivity matrix are extracted from all labels in the cerebral cortex, and the topological features of structural and functional networks in different labels are also extracted.

[0015] Step 180: Based on network topology attributes and brain morphological characteristics, establish feature matrices for patients with brain diseases and healthy subjects, screen out features that are irrelevant to the brain disease rating scale according to relevant analysis methods, and perform feature fusion.

[0016] Step 190: Based on the selected features, train healthy and patient subjects to establish a classification model for brain diseases and a predictive model for the duration of disease.

[0017] Step 200: Repeat steps 180 and 190 until a feature that meets the set requirements is extracted for classification prediction.

[0018] Optionally, the acquired T1 data, dMRI data, and fMRI data are preprocessed using neuroimaging processing software to obtain cortical surface data (pia mater.pial) and fiber tracking data (.trk), specifically including:

[0019] Step 1201: Use FreeSurfer software to perform image preprocessing on T1 data, including head motion correction, intensity normalization, scalp stripping, spatial registration, subcutaneous tissue segmentation, cortical surface construction, surface mapping, cortical segmentation, and obtain pia mater (.pial) cortical surface data.

[0020] Step 1202: Image preprocessing of dMRI data was performed using FSL and DSI-Studio software. First, the dMRI image nodif.nii.gz with magnetic field gradient intensity b=0 was extracted using FSL software and scalp stripping was performed. Then, the generated mask nodif_brain.nii.gz and .bvecs and .bval files were used to perform head motion eddy current correction on the dMRI data. Second, the DSI-Studio software was used to perform deterministic fiber tracing of the whole brain on the corrected dMRI and generate fiber tracing data (.trk).

[0021] Step 1203: Preprocess the fMRI data using FSL software or Matlab toolkit, including time series removal, time slice correction, head motion correction, spatial normalization, spatial smoothing, removal of redundant signals, filtering, and extraction of whole brain time series. Matlab toolkit includes at least one of SPM, DPARSF, and REST.

[0022] Optionally, the transformation matrix M from diffusion magnetic resonance space (dMRI) to structural space (T1) can be obtained using FreeSufer software. d2t The transformation matrix M from functional magnetic resonance imaging (fMRI) to structural T1. f2t The transformation matrix M from volume space to surface space T1st1 ; Obtain the transformation matrix from the MNI space brain partition template to the individual structural space T1, and map the MNI space template to the individual T1 space; Use the individual space brain partition template to remove non-cortical tissue data from T1 cortical spatial fiber tracking data and fMRI data; Specifically, this includes:

[0023] Step 1301: Use affine transformation to obtain the transformation matrix M from diffusion magnetic resonance space (dMRI) to structural space (T1). d2t The transformation matrix M from functional magnetic resonance imaging (fMRI) to structural T1. f2t The transformation matrix M from T1 volume space to cortical space t1 And using the above matrix, T1 cortical spatial fiber tracking data (trk) was obtained. st1 ) and fMRI data ts t1 .

[0024] Specifically, based on the affine transformation method, the transformation matrix M from diffusion magnetic resonance space (dMRI) to structural space (T1) is obtained using the spmregister function in FreeSurfer software. d2t And the transformation matrix M from functional magnetic resonance imaging (fMRI) space (first time point t=1, t∈[1,L], where L represents the number of time points and [1,L] represents integer values ​​from 1 to L) to structural space T1. f2t The transformation matrix M from volume space to surface space is obtained using the mri_info–vox2ras-tkr function. st1 Calculate the spatial coordinates trk of the starting point for tracing each fiber in the T1 cortex. st1 (x,y,z)=M st1 *M d2t *[trk(x,y,z)·(1 / vs trk_x ,1 / vs trk_y ,1 / vs trk_z ),1], where trk(x,y,z) represents the fiber tracking coordinates in the original dMRI space, (·) represents the vector inner product, vs trk_x vs trk_y vs trk_z Let X, Y, and Z represent the voxel sizes in the X, Y, and Z directions of dMRI space, respectively. Calculate the spatial coordinates ts of the T1 cortex at each time point t and for each fMRI data point. st1 (x,y,z)| t＝t =M st1 *M f2t *[ts(x,y,z)·(1 / vsfMRI_x ,1 / vs fMRI_y ,1 / vs fMRI_z ),1], where ts st1 (x,y,z) represents the coordinates of the original fMRI spatial data points at time node t, vs fMRI_x vs fMRI_y vs fMRI_z These represent the voxel sizes in the X, Y, and Z directions of fMRI space, respectively.

[0025] Step 1302: Use nonlinear transformation to obtain the transformation matrix M from the MNI space brain partition template to the individual structural space T1. MNI2t and deformation coefficient W MNI2t And map the Template of the MNI space to the individual T1 space, i.e., Temp ind Using Temp ind Cerebral cortex region mapping, T1 cortex spatial fiber tracking data trk t1 T1 cortical spatial fMRI data ts t1 Remove trk t1 and ts t1 Non-cortical data.

[0026] Specifically, based on the nonlinear transformation method, the transformation matrix M from the MNI space brain partition template to the individual structural space T1 is obtained using the flirt function in the FreeSurfer software. MNI2t The deformation coefficient W is obtained using the fnirt function. MNI2t According to the applywarp function and the transformation matrix M MNI2t Deformation coefficient W MNI2t Map the MNI space Template to the individual T1 space Temp. ind Reading individual spatial templates (Temp) using Matlab software. ind The non-cortical regions are set to 0, and the cortical regions are set to 1. The spatial coordinates trk of the starting point for tracing each fiber in T1 space are calculated. t1 (x,y,z)=M d2t *[trk(x,y,z)·(1 / vs trk_x ,1 / vs trk_y ,1 / vs trk_z ),1], if Temp ind (trk t1 (x),trk t1 (y),trk t1 If (z)) = 0, then the origin or end point of this fiber is outside the cerebral cortex. Remove the corresponding trk.t1 and trk st1 Data. Calculate the spatial coordinates ts of the T1 cortex at each time point t and for each fMRI data point. t1 (x,y,z)| t =M f2t *[ts(x,y,z)·(1 / vs fMRI_x ,1 / vs fMRI_y ,1 / vs fMRI_z ),1], if Temp ind (ts t1 (x),ts t1 (y),ts t1 If (z)) = 0, then the fMRI data point is outside the cerebral cortex. Remove the corresponding ts. t1 and ts st1 data.

[0027] Optionally, calculate the Pearson correlation characterization functional connectivity matrix between microstructure time series, specifically including:

[0028] According to the formula Among them, Act i and Act j They represent the microstructure newΔ i and newΔ j The mean fMRI time-series signal is represented by L, where L represents the time-series length and t represents the time node. A larger correlation coefficient FC(i,j) indicates a higher microstructure newΔ i and newΔ j The stronger the functional connection between them.

[0029] Optionally, calculate the Pearson correlation coefficient between topological features and the SCALE assessment scale, specifically including:

[0030] According to the formula Calculate the Pearson correlation coefficient ρ between topological features and the SCALE assessment scale. X,Y Where X and Y represent topological features and the SCALE assessment scale, respectively, and N represents the number of subjects in the group. and These are the group means for topological characteristics and the SCALE assessment scale, respectively; the correlation coefficient ρ. X,Y The larger the value, the greater the correlation between the topological feature and the assessment scale, and the more effective the feature.

[0031] Secondly, this application provides a device for constructing a brain disease classification model based on multimodal microstructure topological features, the device comprising:

[0032] The first acquisition module is used to acquire the brain disease patient {PAT} data from a pre-set human brain neural image database. n1 and healthy subjects {CON} n2 Multimodal imaging data, including T1-weighted images, dMRI data, and fMRI data; where {…} represents a set; n1 and n2 represent the number of patients with brain diseases and healthy subjects, respectively; a standardized rating scale {SCALE} for assessing symptoms of brain diseases was obtained. n1 、{SCALE} n2 ; Obtain the duration of illness (DUR) in patients with brain diseases n1 ; Obtain brain partition templates at the macroscopic scale of the standard spatial (MNI).

[0033] The first processing module is used to preprocess the acquired T1 data, dMRI data and fMRI data based on neuroimaging processing software, and to obtain cortical surface data (pia mater.pial) and fiber tracking data (.trk).

[0034] The second acquisition module is used to acquire the transformation matrix M from diffusion magnetic resonance space (dMRI) to structural space (T1) using FreeSufer software. d2t The transformation matrix M from functional magnetic resonance imaging (fMRI) to structural T1. f2t The transformation matrix M from volume space to surface space T1 st1 ; Obtain the transformation matrix from the MNI space brain partition template to the individual structural space T1, and map the MNI space template to the individual T1 space; Use the individual space brain partition template to remove non-cortical tissue data from T1 cortical spatial fiber tracking data and fMRI data.

[0035] The second processing module is used to process the individual space template Temp. ind The segmented regions of interest (ROIs) are mapped to the cortical space; the smallest unit of cortical pia mater data is obtained and used as the initial smallest unit (ROI) for microstructural segmentation. Δ And calculate the ROI Δ Area A Δ , establish roi Δ Mapping roi_id to its corresponding label Δ .

[0036] The third processing module is used to downsample and reconstruct the acquired cortical pia mater files, and to use the new triangular mesh units as the final minimum unit (ROI) of the microstructure ROI. newΔ Calculate the ROI of this microstructure newΔArea A newΔ And establish a mapping roi_id between the new microstructure and its corresponding label. newΔ .

[0037] The fourth processing module is used for tracing T1 cortical spatial fiber data (trk). st1 and fMRI data ts st1 Construct a binary structure connectivity matrix SC at the microstructure scale. b Weighted structure connection matrix SC w and functional connection matrix FC.

[0038] The fifth processing module is used for ROI-based... newΔ Mapping roi_id to its corresponding label newΔ Structural connectivity matrix and functional connectivity matrix are extracted from all labels in the cerebral cortex, and the topological features of structural and functional networks in different labels are also extracted.

[0039] The sixth processing module is used to establish feature matrices for patients with brain diseases and healthy subjects based on network topology attributes and brain morphological characteristics, screen out features that are irrelevant to the brain disease rating scale according to relevant analysis methods, and perform feature fusion.

[0040] The seventh processing module is used to train healthy and patient subjects based on selected features to establish a brain disease classification model and a predictive model for the duration of disease.

[0041] The loop module is used to repeat the sixth and seventh processing modules until a feature that meets the set requirements is extracted for classification prediction.

[0042] Optionally, the first processing module specifically includes:

[0043] The first processing submodule is used to perform image preprocessing on T1 data using FreeSurfer software, including head motion correction, intensity normalization, scalp stripping, spatial registration, subcutaneous tissue segmentation, cortical surface construction, surface mapping, cortical segmentation, and acquisition of pia mater (.pial) cortical surface data.

[0044] The second processing submodule is used to preprocess dMRI data using FSL and DSI-Studio software. First, FSL software is used to extract the dMRI image nodif.nii.gz with magnetic field gradient intensity b=0 and perform scalp stripping. Then, the generated mask nodif_brain.nii.gz and .bvecs and .bval files are used to perform head motion eddy current correction on the dMRI data. Second, DSI-Studio software is used to perform deterministic fiber tracing of the whole brain on the corrected dMRI and generate fiber tracing data (.trk).

[0045] The third processing submodule is used to preprocess fMRI data using FSL software or Matlab toolkits, including time series removal, time slice correction, head motion correction, spatial normalization, spatial smoothing, removal of redundant signals, filtering, and extraction of whole-brain time series. The Matlab toolkits include at least one of SPM, DPARSF, and REST.

[0046] Optionally, the second acquisition module specifically includes:

[0047] The first affine processing submodule uses affine transformation to obtain the transformation matrix M from diffuse magnetic resonance space (dMRI) to structural space (T1). d2t The transformation matrix M from functional magnetic resonance imaging (fMRI) to structural T1. f2t The transformation matrix M from T1 volume space to cortical space t1 And using the above matrix, T1 cortical spatial fiber tracking data (trk) was obtained. st1 ) and fMRI data ts t1 Specifically, it includes:

[0048] Based on the affine transformation method, the transformation matrix M from diffuse magnetic resonance space (dMRI) to structural space (T1) is obtained using the spmregister function in FreeSurfer software. d2t And the transformation matrix M from functional magnetic resonance imaging (fMRI) space (first time point t=1, t∈[1,L], where L represents the number of time points and [1,L] represents integer values ​​from 1 to L) to structural space T1. f2t The transformation matrix M from volume space to surface space is obtained using the mri_info–vox2ras-tkr function. st1 ; Calculate the spatial coordinates trk of the starting point for tracing each fiber in the T1 cortex. st1 (x,y,z)=M st1 *M d2t *[trk(x,y,z)·(1 / vs trk_x,1 / vs trk_y ,1 / vs trk_z ),1], where trk(x,y,z) are the fiber tracking coordinates in the original dMRI space, (·) represents the vector inner product, vs trk_x vs trk_y vs trk_z Let X, Y, and Z represent the voxel sizes in the X, Y, and Z directions of dMRI space, respectively; calculate the spatial coordinates ts of the T1 cortex at each time node t and each fMRI data point. st1 (x,y,z)| t＝t =M st1 *M f2t *[ts(x,y,z)·(1 / vs fMRI_x ,1 / vs fMRI_y ,1 / vs fMRI_z ),1], where ts st1 (x,y,z) represents the coordinates of the original fMRI spatial data points at time node t, vs fMRI_x vs fMRI_y vs fMRI_z These represent the voxel sizes in the X, Y, and Z directions of fMRI space, respectively.

[0049] The second nonlinear processing submodule uses nonlinear transformation to obtain the transformation matrix M from the MNI space brain partition template to the individual structural space T1. MNI2t and deformation coefficient W MNI2t And map the Template of the MNI space to the individual T1 space, i.e., Temp ind Using Temp ind Cerebral cortex region mapping, T1 cortex spatial fiber tracking data trk t1 T1 cortical spatial fMRI data ts t1 Remove trk t1 and ts t1 Non-cortical tissue data, specifically including:

[0050] Based on a nonlinear transformation method, the transformation matrix M from the MNI spatial brain partition template to the individual structural space T1 is obtained using the flirt function in the FreeSurfer software. MNI2t The deformation coefficient W is obtained using the fnirt function. MNI2t According to the applywarp function and the transformation matrix M MNI2t Deformation coefficient W MNI2t Map the MNI space Template to the individual T1 space Temp. ind Read the individual spatial template Temp in Matlab software indThe non-cortical regions are set to 0, and the cortical regions are set to 1; the spatial coordinates trk of the starting point for tracing each fiber in T1 space are calculated. t1 (x,y,z)=M d2t *[trk(x,y,z)·(1 / vs trk_x ,1 / vs trk_y ,1 / vs trk_z ),1], if Temp ind (trk t1 (x),trk t1 (y),trk t1 If (z)) = 0, then the origin or end point of this fiber is outside the cerebral cortex. Remove the corresponding trk. t1 and trk st1 Data; calculate the spatial coordinates ts of the T1 cortex at each time point t and each fMRI data point. t1 (x,y,z)| t =M f2t *[ts(x,y,z)·(1 / vs fMRI_x ,1 / vs fMRI_y ,1 / vs fMRI_z ),1], if Temp ind (ts t1 (x),ts t1 (y),ts t1 If (z)) = 0, then the fMRI data point is outside the cerebral cortex. Remove the corresponding ts. t1 and ts st1 data.

[0051] Optionally, the fourth processing module is specifically used for:

[0052] According to the formula Calculate the Pearson correlation characterizing functional connectivity matrix between microstructure time series; where Act i and Act j They represent the microstructure newΔ i and newΔ j The mean fMRI time-series signal is represented by L, where L represents the time-series length and t represents the time node. A larger correlation coefficient FC(i,j) indicates a higher microstructure newΔ i and newΔ j The stronger the functional connection between them.

[0053] Optionally, the sixth processing module is specifically used for:

[0054] According to the formula Calculate the Pearson correlation coefficient ρ between topological features and the SCALE assessment scale. X,YWhere X and Y represent topological features and the SCALE assessment scale, respectively, and N represents the number of subjects in the group. and Represent the group mean of topological characteristics and the SCALE assessment scale, respectively; correlation coefficient ρ X,Y The larger the value, the greater the correlation between the topological feature and the assessment scale, and the more effective the feature.

[0055] Thirdly, this application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the brain disease classification models based on multimodal microstructure topological features of the first aspect. The beneficial effects of the technical solutions provided by the embodiments of this application are:

[0056] This application provides a method and apparatus for constructing a brain disease classification model based on multimodal microstructure topological features. It introduces cortical surface data into the structural space and constructs structural connectivity data and functional connectivity data at the millimeter-scale microstructure, narrowing the calibration range of brain regions. This allows for fine-grained extraction of lesion features deep within brain regions, overcoming the limitations of existing brain disease classification models based on MRI technology and the human brain connectome, which struggle to extract abnormal brain features at the microstructure scale and exhibit inaccurate extraction of lesion regions and features. Furthermore, this application allows for setting a downsampling ratio and selecting the resolution at the microstructure scale and the resolution of the macroscale brain segmentation template. By integrating multi-scale and multi-modal features, a strategy for assessing common and specific damage in diseases can be effectively established, improving the classification accuracy of the model in this application. Before constructing the brain network, the mapping relationship between different modal spaces will be obtained through translation and rotation transformations, thereby constructing the brain network in the individual space and extracting feature information in the individual space, which is beneficial for extracting common and specific features of diseases. In addition, this application selects relatively mature and simple machine learning and deep learning models for classification and prediction, avoiding the problems caused by the large dimension of human brain structure and functional connectivity data at the microstructural scale in practical applications, which is not conducive to feature extraction, feature fusion and model training. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is a flowchart illustrating a method for constructing a brain disease classification model based on multimodal microstructure topological features, as provided in an embodiment of this application.

[0059] Figure 2 yes Figure 1 A schematic diagram of the execution flow for step 120;

[0060] Figure 3 yes Figure 1 A schematic diagram of the execution flow for step 130;

[0061] Figure 4 This is a structural block diagram of a device for constructing a brain disease classification model based on multimodal microstructure topological features, provided in an embodiment of this application.

[0062] Figure 5 This is a structural block diagram of the first processing module 402;

[0063] Figure 6 This is a structural block diagram of the second acquisition module 403. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0065] The terms “first,” “second,” “third,” “fourth,” “fifth,” “sixth,” “seventh,” and “eighth,” etc. (if present), in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein.

[0066] Furthermore, the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, such that a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or apparatus.

[0067] The following will refer to Figures 1-6 As shown, a classification model construction method and apparatus based on multimodal microstructure topological features according to an embodiment of this application will be described in detail.

[0068] refer to Figure 1 As shown in the figure, a method for constructing a brain disease classification model based on multimodal microstructure topological features according to an embodiment of this application includes:

[0069] Step 110: Obtain the brain disease patient {PAT} from a pre-set human brain neuroimaging database. n1 and healthy subjects {CON} n2Multimodal imaging data, including T1-weighted images, dMRI data, and fMRI data; where {…} represents a set; n1 and n2 are the number of patients with brain diseases and healthy subjects, respectively; a standardized rating scale {SCALE} for assessing symptoms of brain diseases was obtained. n1 、{SCALE} n2 ; Obtain the duration of illness (DUR) in patients with brain diseases n1 ; Obtain brain partition templates at the macroscopic scale of the standard spatial (MNI).

[0070] It should be noted that brain diseases can include schizophrenia, Parkinson's disease, Alzheimer's disease and its subtypes, which are brain diseases with unclear exact causes and pathogenesis, and diseases with similar clinical manifestations.

[0071] T1 data is a type of sMRI data that reflects the structure of the human brain.

[0072] dMRI data, including image files (.nii), magnetic field gradient direction files (.bvecs), and magnetic field gradient intensity files (.bval), can characterize the connections and course distribution of nerve fibers in the brain.

[0073] fMRI data, with blood oxygen level dependent (BOLD) signals as the core, are used to detect and locate brain function.

[0074] dMRI data can be acquired using diffusion tensor imaging or diffusion spectroscopy imaging, and this application does not limit the method.

[0075] For example, in order to analyze the differences between healthy subjects and patients with brain diseases, it is required that there are no significant differences in the distribution of gender, age, and education level between the two groups of subjects. The specific statistical test methods for differences will not be elaborated here.

[0076] For example, to analyze the impact of disease duration on brain disease, patients with brain disease will be divided into short-term {DUR1} based on the duration of the disease. n3 and long-term {DUR2} n4 Two groups (n3+n4=n1) are required, and the two groups of subjects are required to have no significant differences in the distribution of gender, age and education level. The specific methods of dividing the two groups and the statistical test methods for differences will not be described in this application.

[0077] It should be noted that the Standardized Assessment Scale for Symptoms of Brain Diseases (SCALE) n1 、{SCALE} n2This system can be used to characterize the severity of symptoms of brain diseases, and also to assess the functional performance of patients with brain diseases and healthy subjects during functional task tests. The embodiments in this application are not specifically limited, and those skilled in the art can refer to existing technologies. The functional task test indicators include test scores after performing cognitive tasks such as emotion processing, executive function, memory tasks, language tasks, and learning tasks. The embodiments in this application are not specifically limited, and those skilled in the art can refer to existing technologies to select a combination of one or more task test scores to create a brain disease rating scale.

[0078] For example, the Standard Minute Area (MNI) macroscale brain partitioning template divides the cerebral cortex into several segmented regions (ROI = {ROI1, ROI2, ... ROI}). i ,…,ROI N} i∈(1,N) (N represents the number of ROIs). The MNI spatial brain partition template can be created using one or more of the following templates: Freesurfer Desikan-Killiany template (N=68), Anotomical AutomaticLabeling template (N=90), Brainnectome template (N=246), and HCP multi-modal template (N=360). This application does not limit this.

[0079] Step 120: Preprocess the acquired T1 data, dMRI data and fMRI data using neuroimaging processing software to obtain cortical surface data (pia mater.pial) and fiber tracking data (.trk).

[0080] Specifically, the execution process of step 120 may include the following sub-steps, which will be discussed in conjunction with the appendix below. Figure 2 The execution process of step 120 is explained in detail.

[0081] Step 1201: Use FreeSurfer software to perform image preprocessing on T1 data, including head motion correction, intensity normalization, scalp stripping, spatial registration, subcutaneous tissue segmentation, cortical surface construction, surface mapping, cortical segmentation, and obtain pia mater (.pial) cortical surface data.

[0082] For example, in order to extract cortical surface features, other cortical surface data can be obtained, such as thickness files (.thickness), curvature files (.curv), etc. This application does not limit the scope of these data.

[0083] It should be noted that cortical segmentation data includes cortical partition data (.annot) and subcutaneous tissue data (Aseg). To ensure the preprocessing results of T1 data, it is necessary to check whether the scalp-peeling images brainmask.mgz or pia mater data.pial of healthy subjects and patients with brain diseases are of good quality. If not, the parameter watershedthreshold needs to be readjusted to correct brainmask.mgz or pia mater.pial needs to be manually corrected based on FreeSurfer to obtain corrected cortical segmentation data. The specific correction methods will not be elaborated here.

[0084] Step 1202: Image preprocessing of dMRI data was performed using FSL and DSI-Studio software. First, the dMRI image nodif.nii.gz with magnetic field gradient intensity b=0 was extracted using FSL software and scalp stripping was performed. Then, the generated mask nodif_brain.nii.gz and .bvecs and .bval files were used to perform head motion eddy current correction on the dMRI data. Second, the DSI-Studio software was used to perform deterministic fiber tracing of the whole brain on the corrected dMRI and generate fiber tracing data (.trk).

[0085] It should be noted that during image preprocessing, the quality of the resulting image nodif_brain.nii.gz needs to be checked. If it is not good, the fractional intensity threshold parameter needs to be readjusted for correction. The specific fiber tracing steps include generating the SRC and performing quality inspection, fiber reconstruction, calculating anisotropy index (FA), mean diffusion index (MD), and other indicators, as well as fiber tracing. The fiber reconstruction method is either DTI or GQI. Fiber tracing termination conditions are set, including fiber count (TractNum), minimum fiber length (MinLength), maximum fiber count (MaxLength), and steering angle (Angular); other parameters such as Smoothing and Seed Orientation are also set.

[0086] Step 1203: Preprocess the fMRI data using FSL software or Matlab toolkit, including time series removal, time slice correction, head motion correction, spatial normalization, spatial smoothing, removal of redundant signals, filtering, and extraction of whole brain time series. Matlab toolkit includes at least one of SPM, DPARSF, and REST, and can be any one or more of them. This application embodiment does not make specific limitations, and those skilled in the art can refer to the prior art for selection.

[0087] Step 130: Use FreeSufer software to obtain the transformation matrix M from diffusion magnetic resonance space (dMRI) to structural space (T1). d2t The transformation matrix M from functional magnetic resonance imaging (fMRI) to structural T1. f2t The transformation matrix M from volume space to surface space T1 st1 ; Obtain the transformation matrix from the MNI space brain partition template to the individual structural space T1, and map the MNI space template to the individual T1 space; Use the individual space brain partition template to remove non-cortical tissue data from T1 cortical spatial fiber tracking data and fMRI data.

[0088] Specifically, the execution process of step 130 may include the following sub-steps, which will be discussed in conjunction with the appendix below. Figure 3 The execution process of step 130 is explained in detail.

[0089] Step 1301: Use affine transformation to obtain the transformation matrix M from diffusion magnetic resonance space (dMRI) to structural space (T1). d2t The transformation matrix M from functional magnetic resonance imaging (fMRI) to structural T1. f2t The transformation matrix M from T1 volume space to cortical space t1 And using the above matrix, T1 cortical spatial fiber tracking data (trk) was obtained. st1 ) and fMRI data ts t1 .

[0090] Specifically, based on the affine transformation method, the transformation matrix M from diffusion magnetic resonance space (dMRI) to structural space (T1) is obtained using the spmregister function in FreeSurfer software. d2t And the transformation matrix M from functional magnetic resonance imaging (fMRI) space (first time point t=1, t∈[1,L], where L represents the number of time points) to structural space T1. f2t The transformation matrix M from volume space to surface space is obtained using the mri_info–vox2ras-tkr function. st1 Calculate the spatial coordinates trk of the starting point for tracing each fiber in the T1 cortex. st1 (x,y,z)=M st1 *M d2t *[trk(x,y,z)·(1 / vs trk_x ,1 / vs trk_y ,1 / vs trk_z),1], where trk(x,y,z) are the fiber tracking coordinates in the original dMRI space, (·) represents the vector inner product, vs trk_x vs trk_y vs trk_z Let X, Y, and Z represent the voxel sizes in the X, Y, and Z directions of dMRI space, respectively. Calculate the spatial coordinates ts of the T1 cortex at each time point t and for each fMRI data point. st1 (x,y,z)| t＝t =M st1 *M f2t *[ts(x,y,z)·(1 / vs fMRI_x ,1 / vs fMRI_y ,1 / vs fMRI_z ),1], where ts st1 (x, y, z) represents the coordinates of the original fMRI spatial data points at time node t, vs fMRI_x vs fMRI_y vs fMRI_z These represent the voxel sizes in the X, Y, and Z directions of fMRI space, respectively.

[0091] Step 1302: Use nonlinear transformation to obtain the transformation matrix M from the MNI space brain partition template to the individual structural space T1. MNI2t and deformation coefficient W MNI2t And map the Template of the MNI space to the individual T1 space, i.e., Temp ind Using Temp ind Cerebral cortex region mapping, T1 cortex spatial fiber tracking data trk t1 T1 cortical spatial fMRI data ts t1 Remove trk t1 and ts t1 Non-cortical data.

[0092] Specifically, based on the nonlinear transformation method, the transformation matrix M from the MNI space brain partition template to the individual structural space T1 is obtained using the flirt function in the FreeSurfer software. MNI2t The deformation coefficient W is obtained using the fnirt function. MNI2t According to the applywarp function and the transformation matrix M MNI2t Deformation coefficient W MNI2t Map the MNI space Template to the individual T1 space Temp. ind Reading individual spatial templates (Temp) using Matlab software. indThe non-cortical regions are set to 0, and the cortical regions are set to 1. The spatial coordinates trk of the starting point for tracing each fiber in T1 space are calculated. t1 (x,y,z)=M d2t *[trk(x,y,z)·(1 / vs trk_x ,1 / vs trk_y ,1 / vs trk_z ),1], if Temp ind (trk t1 (x),trk t1 (y),trk t1 If (z)) = 0, then the origin or end point of this fiber is outside the cerebral cortex. Remove the corresponding trk. t1 and trk st1 Data. Calculate the spatial coordinates ts of the T1 cortex at each time point t and for each fMRI data point. t1 (x,y,z)| t =M f2t *[ts(x,y,z)·(1 / vs fMRI_x ,1 / vs fMRI_y ,1 / vs fMRI_z ),1], if Temp ind (ts t1 (x),ts t1 (y),ts t1 If (z)) = 0, then the fMRI data point is outside the cerebral cortex. Remove the corresponding ts. t1 and ts st1 data.

[0093] Step 140: Create an individual space template (Temp). ind The segmented regions of interest (ROIs) are mapped to the cortical space; the smallest unit of cortical pia mater data is obtained and used as the initial smallest unit (ROI) for microstructural segmentation. Δ And calculate the ROI Δ Area A Δ , establish roi Δ Mapping roi_id to its corresponding label Δ .

[0094] Specifically, the `mri_annotation2label` function in the FreeSurfer software is used to map each segmented region (ROI) to the cortical space (.label file, label = {label1, label2, ... label...). i ,…,label N} i∈NThe pia mater data (.pial) is read in Matlab, and the smallest unit formed by the cortical surface—a triangular mesh—is generated. This triangular mesh is represented by triangles composed of vertices and faces. The specific method for reading the pia mater data in Matlab will not be elaborated upon here. A roi is then established. Δ Mapping roi_id to its corresponding label Δ Then a macroscopic label corresponds to multiple microstructural units (ROIs). Δ The method used is as follows: if the roi is composed of Δ At least two of the three points are in the label. i If it is inside, then the ROI is determined. Δ Belongs to label i ;

[0095] Step 150: Downsample and reconstruct the acquired cortical pia mater file, and use the new triangular mesh units as the final minimum unit (ROI) of the microstructure ROI. newΔ Calculate the ROI of this microstructure newΔ Area A newΔ And establish a mapping roi_id between the new microstructure and its corresponding label. newΔ .

[0096] It should be noted that the iso2mesh toolkit in Matlab was used to downsample and reconstruct the cortical pia mater data. The downsampling ratio was set to 'rate', and to reduce computational complexity while preserving microstructural patterns, the selected downsampling ratio 'rate' was between 0.01 and 0.1. After downsampling, the roi of the neocortical surface forming units need to be reread. newΔ , and its points vertices new and faces new and establish roi newΔ with roi Δ The mapping relationship, and roi newΔ Mapping roi_id to its corresponding label newΔ The method used is ROI newΔ Center point and roi Δ Minimum Euclidean distance of the center point, roi newΔ The label and the nearest ROI Δ They belong to the same label.

[0097] Step 160: Based on T1 cortical spatial fiber tracking data trk st1 and fMRI data ts st1 Construct a binary structure connectivity matrix SC at the microstructure scale. b Weighted structure connection matrix SCw and functional connection matrix FC.

[0098] Specifically, based on T1 cortical spatial fiber tracking data trk st1 Establish the starting and ending coordinates of each fiber tracking point and ROI. newΔ The mapping of the coordinates and the mapping of the corresponding labels are based on the fiber tracking coordinates trk. st1 (x,y,z) and roi newΔ The minimum Euclidean distance between the center points, trk st1 The label and ROI of (x,y,z) newΔ Those with the same label are used to construct a binary structured connection matrix SC. b and weighted structure connection matrix SC w , of which SC b (i,j)=1,i∈(1,N) represents the microstructure newΔ i With newΔ j Fiber connections exist between structures, indicating structural connections; conversely, they do not exist if no fiber connections are present. Weighted structural connection matrix. Where hSC(i,j) represents the microstructure newΔ i With newΔ j The total number of fibers present. The area of ​​region i, representing microstructure, is used to remove the influence of cortical surface area on structural connectivity. The weighted structural connectivity matrix can also be constructed by normalizing fiber quantity, average fiber length, etc. This application does not impose specific limitations on the embodiments; those skilled in the art can refer to existing technologies for selection.

[0099] Based on T1 cortical spatial fMRI data ts st1 Establish the spatial coordinates of each fMRI data point and its ROI. newΔ The mapping of fMRI and its associated label is based on the fMRI coordinates ts at time t=1. st1 (x,y,z) and roi newΔ The minimum Euclidean distance between the center points, ts st1 The label and ROI of (x,y,z) newΔ Given the same label, extract the fMRI time series ts for each coordinate (x, y, z). st1 (x,y,z,t), then a roi newΔ This may correspond to multiple fMRI coordinates and multiple time series. Calculate the microstructure newΔ. i The mean of all fMRI time series within the time frame was used as the time series of the microstructure, and the newΔ value of the microstructure was calculated using the Pearson correlation coefficient. i With newΔ jThe degree of correlation between them is characterized by the functional connectivity matrix FC(i,j). The functional connectivity matrix can also be constructed by taking the absolute value of the Pearson correlation coefficient, taking the partial correlation coefficient, etc. The implementation of this application is not specifically limited, and those skilled in the art can refer to the prior art for selection.

[0100] It should be noted that, according to the formula Calculate the Pearson correlation characterizing functional connectivity matrix between microstructure time series; where Act i and Act j They represent the microstructure newΔ i and newΔ j The mean fMRI time-series signal is represented by L, where L represents the time-series length and t represents the time node. A larger correlation coefficient FC(i,j) indicates a higher microstructure newΔ i and newΔ j The stronger the functional connection between them.

[0101] Step 170: Based on ROI newΔ Mapping roi_id to its corresponding label newΔ Structural connectivity matrix and functional connectivity matrix are extracted from all labels in the cerebral cortex, and the topological features of structural and functional networks in different labels are also extracted.

[0102] Specifically, according to the formula Obtain the binary structure connection matrix label(SC) within a specific label. b ), weighted structure connection matrix label(SC) w The network topology features are defined as follows: (a matrix of matrices) and (b matrix of functional connectivity labels (FC)). These three matrices record the microstructural connectivity information within each ROI at a macroscopic scale. It should be noted that the label can be a disease-related brain region, such as the prefrontal cortex, or a combination of one or more other brain regions. This application does not impose specific limitations on the chosen label; those skilled in the art can refer to existing technologies for selection. The network topology features can be selected as a combination of one or more of the following: average node degree, average clustering coefficient, average feature path length, small-world attribute, modularity, and community attribute. This application does not impose specific limitations on the chosen network topology features; those skilled in the art can refer to existing technologies for selection. Specific methods for extracting network topology features will not be elaborated upon here.

[0103] Step 180: Based on network topology attributes and brain morphological characteristics, establish feature matrices for patients with brain diseases and healthy subjects, screen out features that are irrelevant to the brain disease rating scale according to relevant analysis methods, and perform feature fusion.

[0104] Specifically, Feature pat (lab,f,s) represents the feature matrix of patients with brain diseases, where lab represents the index of the label, f represents the index of the multimodal topological feature, and s represents the index of the patient. con (lab,f,s) represents the feature matrix of the healthy subject. This feature matrix not only contains features of multiple modalities and labels, but also their morphological features, including the average cortical thickness, average cortical curvature, average anisotropy index, and average diffusion under each label. The embodiments in this application do not impose specific limitations, and those skilled in the art can refer to the prior art to select one or more combinations.

[0105] It should be noted that, in order to analyze the impact of disease duration on brain diseases, brain diseases are divided into short-term {DUR1} n3 and long-term {DUR2} n4 Two groups, among which the feature matrix of patients with short-term brain disease is Feature DUR1 (lab,f,s), the feature matrix of patients with long-term brain diseases is Feature DUR2 (lab,f,s). In conducting inter-group difference analysis, to avoid the influence of gender, age, and education level, linear regression was used to regress the above covariates. The specific regression process will not be described in detail here.

[0106] Specifically, according to the formula Calculate the Pearson correlation coefficient ρ between topological features and the SCALE assessment scale. X,Y Feature selection is performed, where X and Y represent topological features and the SCALE assessment scale, respectively, and N represents the number of subjects in the group. and These represent the group means for topological characteristics and the SCALE assessment scale, respectively. Correlation coefficient ρ X,Y The larger the correlation coefficient, the greater the correlation between the topological feature and the evaluation scale, and thus the more effective the feature. This application does not specify a particular threshold for the correlation coefficient used in feature selection; those skilled in the art can refer to existing technologies for selection.

[0107] After feature selection, the effective features need to be sorted and fused. Feature fusion methods include linear fusion, nonlinear fusion, dictionary learning, etc. The embodiments of this application do not limit the specific methods, and those skilled in the art can refer to the prior art for selection.

[0108] Step 190: Based on the selected features, train healthy and patient subjects to establish a classification model for brain diseases and a predictive model for the duration of disease.

[0109] For example, in group analyses of healthy subjects and patients with brain diseases, to prevent overfitting and obtain a relatively stable classification model, cross-validation is used. The data is randomly divided into training, test, and validation sets, and the experiment is repeated multiple times until a good training model is obtained. Similarly, in group analyses of patients with short-term and long-term brain diseases, cross-validation is used, the data is randomly divided into training, test, and validation sets, and the experiment is repeated multiple times until a good training model is obtained.

[0110] It should be noted that the machine learning algorithms used in the model include, but are not limited to, nearest neighbors, random forests, decision trees, support vector machines, clustering, neural networks, and deep learning. This application does not impose specific limitations on these algorithms, and those skilled in the art can refer to existing technologies to select one or more for model evaluation. The cross-validation methods used include, but are not limited to, holdout cross-validation, k-fold cross-validation, and leave-one-out cross-validation. This application does not impose specific limitations on these methods, and those skilled in the art can refer to existing technologies to select one or more for model evaluation.

[0111] Step 200: Repeat steps 180 and 190 until a feature that meets the set requirements is extracted for classification prediction.

[0112] It should be noted that the required value can be set according to the current accuracy of brain disease classification, generally between 0.75 and 1. For example, it can be any value between 0.75, 0.8, 0.85, 0.9, 0.95, 1 or 0.75-1. This application embodiment does not impose any restrictions.

[0113] This application provides a method for constructing a brain disease classification model based on multimodal microstructure topological features. It introduces cortical surface data into the structural space and constructs structural connectivity data and functional connectivity data at the millimeter-level microstructure scale. This narrows the calibration range of brain regions, enabling fine-grained extraction of lesion features from deep within the brain. This overcomes the limitations of existing brain disease classification models based on MRI technology and the human brain connectome, which struggle to extract abnormal brain features at the microstructure scale and suffer from inaccurate extraction of lesion regions and features. Furthermore, this application allows setting a downsampling ratio and selecting the resolution at the microstructure scale and the resolution of the macroscale brain segmentation template, thus integrating... By combining multi-scale and multi-modal features, a strategy for assessing common and specific damage in diseases can be effectively established, improving the classification accuracy of the model in this application. Before constructing the brain network, the mapping relationship between different modal spaces will be obtained through translation and rotation transformations, thereby constructing the brain network in the individual space and extracting feature information in the individual space, which is beneficial for extracting common and specific features of diseases. In addition, this application selects relatively mature and simple machine learning and deep learning models for classification and prediction, avoiding the problems caused by the large dimension of human brain structure and functional connectivity data at the microstructural scale in practical applications, which is not conducive to feature extraction, feature fusion and model training.

[0114] Furthermore, this application provides a method for constructing a brain disease classification model based on multimodal microstructure topological features. This method overcomes the limitations of existing brain disease classification models based on MRI technology and the human brain connectome, which can only rely on a large range of brain regions and the connectivity between regions, and cannot delve into the fine-grained features of lesions within brain regions. It not only effectively extracts structural morphological features and network topological features at the individualized brain microstructure scale, but also extracts functional network topological features, and performs multimodal, multi-scale, and multi-feature fusion. This allows for fine-grained analysis of brain abnormalities from multiple perspectives, extracting disease-specific and common indicators or features, and avoiding poor classification results due to a lack of individualized local microstructure information. In disease classification based on extracted features, this application can train and classify diseases using a variety of selectable machine learning and deep learning models, selecting the most advantageous features and models to reveal the structure, function, and pathological mechanisms of the diseased brain, establish assessment strategies for common and specific impairments in brain diseases, and improve the model's classification performance.

[0115] Figure 4 This application provides an apparatus for constructing a brain disease classification model based on multimodal microstructure topological features, with reference to... Figure 4 As shown, the brain disease classification device includes:

[0116] The first acquisition module 401 is used to acquire the brain disease patient {PAT} from a pre-set human brain neural image database. n1and healthy subjects {CON} n2 Multimodal imaging data, including T1-weighted images, dMRI data, and fMRI data. Where {…} represents a set; n1 and n2 represent the number of patients with brain diseases and healthy subjects, respectively; a standardized rating scale {SCALE} for assessing symptoms of brain diseases was obtained. n1 、{SCALE} n2 ; Obtain the duration of illness (DUR) in patients with brain diseases n1 ; Obtain brain partition templates at the macroscopic scale of the standard spatial (MNI).

[0117] The first processing module 402 is used to preprocess the acquired T1 data, dMRI data and fMRI data based on neuroimaging processing software, and to obtain cortical surface data (pia mater.pial) and fiber tracking data (.trk).

[0118] Optionally, the first processing module 402 specifically includes:

[0119] The first processing submodule 4021 is used to perform image preprocessing on T1 data using FreeSurfer software, including head motion correction, intensity normalization, scalp stripping, spatial registration, subcutaneous tissue segmentation, cortical surface construction, surface mapping, cortical segmentation, and to obtain pia mater (.pial) cortical surface data.

[0120] The second processing submodule 4022 is used to perform image preprocessing on dMRI data using FSL software and DSI-Studio software. First, FSL software is used to extract the dMRI image nodif.nii.gz with magnetic field gradient intensity b=0 and perform scalp stripping operation. Then, the generated mask nodif_brain.nii.gz and .bvecs and .bval files are used to perform head motion eddy current correction on the dMRI data. Second, DSI-Studio software is used to perform deterministic fiber tracing of the whole brain on the corrected dMRI and generate fiber tracing data (.trk).

[0121] The third processing submodule 4023 is used to preprocess fMRI data using FSL software or Matlab toolkit, including time series removal, time slice correction, head motion correction, spatial normalization, spatial smoothing, removal of redundant signals, filtering, and extraction of whole brain time series. Matlab toolkit includes at least one of SPM, DPARSF, and REST. This application embodiment does not make specific limitations, and those skilled in the art can refer to the prior art for selection.

[0122] The second acquisition module 403 is used to acquire the transformation matrix M from diffusion magnetic resonance space dMRI to structural space T1 using FreeSufer software.d2t The transformation matrix M from functional magnetic resonance imaging (fMRI) to structural T1. f2t The transformation matrix M from volume space to surface space T1 st1 ; Obtain the transformation matrix from the MNI space brain partition template to the individual structural space T1, and map the MNI space template to the individual T1 space; Use the individual space brain partition template to remove non-cortical tissues from T1 cortical spatial fiber tracking data and fMRI data.

[0123] Optionally, the second acquisition module 403 specifically includes:

[0124] The first affine processing submodule 4031 uses affine transformation to obtain the transformation matrix M from diffuse magnetic resonance space (dMRI) to structural space (T1). d2t The transformation matrix M from functional magnetic resonance imaging (fMRI) to structural T1. f2t The transformation matrix M from T1 volume space to cortical space t1 And using the above matrix, T1 cortical spatial fiber tracking data (trk) was obtained. st1 ) and fMRI data ts t1 .

[0125] The second nonlinear processing submodule 4032 uses nonlinear transformation to obtain the transformation matrix M from the MNI space brain partition template to the individual structural space T1. MNI2t and deformation coefficient W MNI2t And map the Template of the MNI space to the individual T1 space, i.e., Temp ind Using Temp ind Cerebral cortex region mapping, T1 cortex spatial fiber tracking data trk t1 T1 cortical spatial fMRI data ts t1 Remove trk t1 and ts t1 Non-cortical data.

[0126] The second processing module 404 is used to process the individual space template Temp. ind The segmented regions of interest (ROIs) are mapped to the cortical space; the smallest unit of cortical pia mater data is obtained and used as the initial smallest unit (ROI) for microstructural segmentation. Δ And calculate the ROI Δ Area A Δ , establish roi Δ Mapping roi_id to its corresponding label Δ .

[0127] The third processing module 405 is used to downsample and reconstruct the acquired cortical pia mater file, and use the new triangular mesh unit as the final minimum unit roi of the microstructure ROI. newΔ Calculate the ROI of this microstructure newΔ Area A newΔ And establish a mapping roi_id between the new microstructure and its corresponding label. newΔ .

[0128] The fourth processing module 406 is used for processing T1 cortical spatial fiber tracking data trk st1 and fMRI data ts st1 Construct a binary structure connectivity matrix SC at the microstructure scale. b Weighted structure connection matrix SC w and functional connection matrix FC.

[0129] Optionally, the fourth processing module 406 is specifically used for:

[0130] According to the formula Calculate the Pearson correlation characterizing functional connectivity matrix between microstructure time series; where Act i and Act j They represent the microstructure newΔ i and newΔ j The mean fMRI time-series signal is represented by L, where L represents the time-series length and t represents the time node. A larger correlation coefficient FC(i,j) indicates a higher microstructure newΔ i and newΔ j The stronger the functional connection between them.

[0131] The fifth processing module 407 is used for ROI-based processing. newΔ Mapping roi_id to its corresponding label newΔ Structural connectivity matrix and functional connectivity matrix are extracted from all labels in the cerebral cortex, and the topological features of structural and functional networks in different labels are also extracted.

[0132] The sixth processing module 408 establishes a feature matrix for patients with brain diseases and healthy subjects based on network topology attributes and brain morphological characteristics. It then filters out features that are irrelevant to the brain disease rating scale according to relevant analysis methods and performs feature fusion.

[0133] Optionally, the sixth processing module 408 is specifically used for:

[0134] According to the formula Calculate the Pearson correlation coefficient ρ between topological features and the SCALE assessment scale.X,Y Feature selection is performed, where X and Y represent topological features and the SCALE assessment scale, respectively, and N represents the number of subjects in the group. and These are the group means for topological characteristics and the SCALE assessment scale, respectively; the correlation coefficient ρ. X,Y The larger the value, the greater the correlation between the topological feature and the assessment scale, and the more effective the feature.

[0135] The seventh processing module 409 is used to train healthy and patient subjects based on selected features to establish a classification model of brain diseases and a predictive model of the duration of disease.

[0136] The loop module 410 is used to repeat steps 6 and 7 until a feature that meets the set requirements is extracted for classification prediction.

[0137] It should be noted that the required value can be set according to the current accuracy of brain disease classification, generally between 0.75 and 1. For example, it can be any value between 0.75, 0.8, 0.85, 0.9, 0.95, 1 or 0.75-1. This application embodiment does not impose any restrictions.

[0138] It should also be noted that the above-described embodiment of the brain disease classification model construction device based on multimodal microstructure topological features is only illustrated by the division of the functional modules described above when classifying brain diseases. In practical applications, the functions described above can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. Furthermore, the above-described embodiment of the brain disease classification model construction device based on multimodal microstructure topological features and the embodiment of the method for constructing a brain disease classification model based on multimodal microstructure topological features belong to the same concept, and their specific implementation process is detailed in the method embodiment, which will not be repeated here.

[0139] In addition, this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the above-described methods for constructing a brain disease classification model based on multimodal microstructure topological features.

[0140] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0141] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0142] The above embodiments are merely preferred embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for constructing a brain disease classification model based on multi-modal microstructure topological features, characterized in that, Comprising the following steps: Step 110: obtaining multi-modal image data of brain disease patients and healthy subjects from a pre-set human brain nerve image database, including T1, dMRI data and fMRI data; obtaining a standardized assessment scale for assessing brain disease symptoms and the disease duration of the patients; obtaining a standard space macro-scale brain partition template; Step 120: pre-processing the obtained T1 data, dMRI data and fMRI data based on a neural image processing software, and obtaining cortical surface data and fiber tracking data; Step 130: obtaining a conversion matrix from diffusion magnetic resonance space dMRI to structural space T1, a conversion matrix from functional magnetic resonance space fMRI to structural space T1, and a conversion matrix from T1 body space to cortical space; obtaining a conversion matrix from MNI space brain partition template Template to individual structural space T1, and mapping the Template in MNI space to individual T1 space; Using the individual space brain partition template to remove non-brain cortical tissue data in the T1 cortical space fiber tracking data and the fMRI data; Step 140: mapping the segmented region ROI of the individual space template to the cortical space; Obtaining the smallest unit of the cortical pia mater data as the initialization smallest unit of the microstructure segmentation region, and calculating the area of the initialization smallest unit, and establishing the mapping of the microstructure and the corresponding label; Step 150: downsampling and reconstructing the obtained cortical pia mater file, and taking the new triangular mesh unit as the final smallest unit of the microstructure ROI, calculating the area of the final smallest unit, and establishing the mapping of the new microstructure and the corresponding label; Step 160: based on the T1 cortical space fiber tracking data and the fMRI data, constructing a binary structure connection matrix, a weighted structure connection matrix and a functional connection matrix of the microstructure scale; Step 170: based on the mapping of the microstructure and the corresponding label, the structure connection matrix and the functional connection matrix; Respectively extracting the structure connection matrix and the functional connection matrix in all labels of the brain cortex, and extracting the topological features of the structure network and the functional network in different labels; Step 180: based on the network topological properties and the brain morphological features, establishing a feature matrix of the brain disease patients and the healthy subjects, screening out the features irrelevant to the brain disease according to a correlation analysis method, and performing feature fusion; Step 190: based on the selected features, training the healthy and patient subjects to establish a classification model of the brain disease and a prediction model of the disease duration; Step 200: repeating steps 180 and 190 until a feature meeting the set requirements is extracted for classification and prediction.

2. The method of claim 1, wherein the method is characterized by: The obtained T1 data, dMRI data and fMRI data are pre-processed based on a neural image processing software, and cortical surface data and fiber tracking data are obtained, specifically including: Step 1201: image preprocessing of T1 data; including head motion correction, intensity normalization, scalp stripping, spatial registration, subcutaneous tissue segmentation, cortical surface construction, curved surface mapping, cortical segmentation, and obtaining pia mater cortical surface data; Step 1202: image preprocessing of dMRI data; including extracting dMRI image nodif.nii.gz with magnetic field gradient strength b=0 and performing scalp stripping operation, using generated mask nodif_brain.nii.gz and.bvecs,.bval files to perform head motion vortex correction on dMRI data, and generating fiber tracking data after whole brain deterministic fiber tracking on the corrected dMRI; Step 1203: preprocessing of fMRI data; including removing time series, time layer correction, head motion correction, spatial standardization, spatial smoothing, removing redundant signals, filtering, and extracting whole brain time series.

3. The method of claim 1 or 2, wherein the method is characterized by, Respectively obtain the conversion matrix from the diffusion magnetic resonance space dMRI to the structure space T1, the conversion matrix from the functional magnetic resonance space fMRI to the structure space T1, and the conversion matrix from the T1 body space to the cortical space; obtain the conversion matrix from the MNI space brain partition template Template to the individual structure space T1, and map the Template in the MNI space to the individual T1 space; Use the brain partition template in the individual space to remove non-brain cortical tissue data in the T1 cortical space fiber tracking data and fMRI data, specifically including: Step 1301: using affine transformation to obtain the conversion matrix from the diffusion magnetic resonance space dMRI to the anatomical space T1, the conversion matrix from the functional magnetic resonance space fMRI to the anatomical space T1, and the conversion matrix from the T1 body space to the cortical space, and using the above matrix to obtain the T1 cortical space fiber tracking data and fMRI data; Step 1302: using nonlinear transformation to obtain the conversion matrix from the MNI space brain partition template Template to the individual anatomical space T1 and the deformation coefficient, and mapping the Template in the MNI space to the individual T1 space; using the brain cortical partition demarcation of the individual space template Template, the T1 cortical space fiber tracking data and the T1 cortical space fMRI data, removing non-brain cortical data in the fiber tracking data and the fMRI data.

4. A device for constructing a brain disease classification model based on multi-modal microstructure topological features, characterized in that, It includes: A first acquisition module; For acquiring multi-modal image data of brain disease patients and healthy subjects from a pre-set human brain neural image database, including T1, dMRI data and fMRI data; acquiring a standardized assessment scale for assessing brain disease symptoms and the patient's disease duration; acquiring a standard space macro-scale brain partition template; A first processing module: for preprocessing the acquired T1 data, dMRI data and fMRI data based on a neural image processing software, and obtaining cortical surface data and fiber tracking data; A second acquisition module; The first processing module comprises: The first processing submodule is used for image preprocessing of T1 data; including head motion correction, intensity standardization, head skin peeling, spatial registration, subcutaneous tissue segmentation, cortical surface construction, curved surface mapping, cortex segmentation, and obtaining pia mater cortical surface data; The second processing submodule is used for image preprocessing of dMRI data; including extracting dMRI image nodif.nii.gz with magnetic field gradient strength b=0 and performing head skin peeling operation, using the generated mask nodif_brain.nii.gz and.bvecs,.bval files to perform head motion vortex correction on dMRI data, performing whole brain deterministic fiber tracking on the corrected dMRI, and generating fiber tracking data; The second processing module comprises: The third processing submodule is used for mapping the segmented region ROI of the individual space template to the cortical space; obtaining the smallest unit of the cortical pia mater data and taking it as the initial smallest unit of the microstructure segmented region, calculating the area of the initial smallest unit, and establishing the mapping of the microstructure and the corresponding label; The third processing module comprises: The fourth processing submodule is used for downsampling and reconstructing the obtained cortical pia mater file, taking the new triangular mesh unit as the final smallest unit of the microstructure ROI, calculating the area of the final smallest unit, and establishing the mapping of the new microstructure and the corresponding label; The fourth processing module comprises: The fifth processing submodule is used for constructing a binary structure connection matrix, a weighted structure connection matrix and a functional connection matrix of the microstructure scale based on the T1 cortical space fiber tracking data and the fMRI data; The fifth processing module comprises: The sixth processing submodule is used for extracting the structure connection matrix and the functional connection matrix within all labels of the cerebral cortex based on the mapping of the microstructure and the corresponding label, the structure connection matrix and the functional connection matrix; and extracting the topological features of the structure network and the functional network within different labels; The sixth processing module comprises: The seventh processing submodule is used for training the healthy and patient subjects based on the selected features, establishing a classification model of brain diseases and a prediction model of the length of the disease duration; The loop module comprises: The loop module is used for repeating the sixth processing module and the seventh processing module until a feature meeting the set requirements is extracted for classification and prediction.

5. The device for constructing a brain disease classification model based on a multi-modal microstructure topology feature according to claim 4, characterized in that, The first processing module comprises: The first processing submodule is used for image preprocessing of T1 data; including head motion correction, intensity standardization, head skin peeling, spatial registration, subcutaneous tissue segmentation, cortical surface construction, curved surface mapping, cortex segmentation, and obtaining pia mater cortical surface data; The second processing submodule is used for image preprocessing of dMRI data; including extracting dMRI image nodif.nii.gz with magnetic field gradient strength b=0 and performing head skin peeling operation, using the generated mask nodif_brain.nii.gz and.bvecs,.bval files to perform head motion vortex correction on dMRI data, performing whole brain deterministic fiber tracking on the corrected dMRI, and generating fiber tracking data; A third processing submodule is configured to preprocess the fMRI data, including removing time series, time layer correction, head motion correction, spatial standardization, spatial smoothing, removing redundant signals, filtering, and extracting whole brain time series.

6. The device for constructing a brain disease classification model based on a multi-modal microstructure topology feature according to claim 4 or 5, characterized in that, The second processing module includes: A first affine processing submodule is configured to obtain a conversion matrix from diffusion magnetic resonance space dMRI to anatomical space T1, a conversion matrix from functional magnetic resonance space fMRI to anatomical space T1, and a conversion matrix from T1 body space to cortex space by using affine transformation, and obtain T1 cortex space fiber tracking data and fMRI data by using the above matrices. A second nonlinear processing submodule is configured to obtain a conversion matrix from MNI space brain partition template Template to individual anatomical space T1 and a deformation coefficient by using nonlinear transformation, and map the Template in MNI space to individual T1 space; and remove non-brain cortex data in the fiber tracking data and the fMRI data by using brain cortex partition demarcation of the individual space Template, the T1 cortex space fiber tracking data, and the T1 cortex space fMRI data.

Images