A method for neural dissection heterogeneity analysis integrating standard models and non-negative matrix factorization

Abstract

The application discloses a neural anatomical heterogeneity analysis method integrating standard model and non-negative matrix factorization, and belongs to the field of biomedical signal processing. The application comprises: preprocessing structural magnetic resonance imaging data of patients and corresponding healthy control groups to extract brain morphological indexes; establishing a brain morphological standard model of the patient through a transfer pre-training model to estimate a deviation score matrix of the patient's cortical thickness and surface area, and extracting a non-positive part to obtain a brain morphological atrophy deviation matrix; performing non-negative matrix factorization on the atrophy deviation matrix to generate a preliminary disease factor matrix and a weight composition matrix; and rearranging and sampling the deviation matrix, calculating an average half-decomposition stability coefficient and a reconstruction error to determine the disease factor and the weight composition under the optimal component number. The method of the application integrates spatial pattern information at the group level while retaining individual heterogeneity information, and provides a new technical means for precise diagnosis of neuropsychiatric diseases.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of biomedical image signal processing, and particularly relates to a neural anatomical heterogeneity analysis method integrating a normative model and non-negative matrix factorization. BACKGROUND

[0002] The brain structure of patients with neuropsychiatric diseases shows significant individual heterogeneity, mainly in clinical, pathological, environmental and other aspects. At the population level, due to the superposition of individual symptoms, there is often a large difference between the disease symptoms at the population level and a certain clinical individual. Therefore, how to accurately depict the heterogeneity phenotype of the disease at the individual level is particularly important.

[0003] Brain morphology research can track the development process of a healthy brain and identify potential deviations related to diseases. Brain morphology research uses a normative model (NM) framework to estimate individual differences in the life cycle using large-scale and diverse data sets, providing an effective tool for transferring inferences from the population level (such as case-control studies) to the individual level. The normative model can quantify the differences between individuals and the reference population, and has been widely used in the study of brain structure variation patterns in neurological and neuropsychiatric diseases. However, the deviation scores obtained by the normative model are usually high-dimensional data of multiple brain regions, and existing processing methods often only focus on a single brain region, failing to effectively integrate the spatial distribution relationship between brain regions. This processing method fails to effectively integrate the spatial pattern information at the population level while preserving individual heterogeneity information, resulting in low interpretability of individual heterogeneity results. In addition, the integration of different structural indicators within the same brain region has not been fully appreciated.

[0004] Non-negative matrix factorization (NMF) is an unsupervised multivariate analysis method that effectively extracts spatial patterns in data by decomposing original data into non-negative low-rank matrices, and captures local and global variations in brain cortical structure. NMF can be used to identify spatial patterns between different regions of the brain, especially in multiple morphological measurements (such as brain cortical thickness, surface area, volume, etc.), it can capture the variation rules of different brain regions. NMF has high interpretability and scalability across datasets in neuroimaging data analysis, so it can be applied to the processing of high-dimensional data of the normative model, solving the interpretability problem of existing methods and further revealing abnormal brain structure patterns, providing strong support for individualized diagnosis and treatment. SUMMARY

[0005] The application aims to solve the defects of the prior art, and provides a neural anatomical heterogeneity analysis method integrating a standard model and non-negative matrix factorization, so as to provide a new data analysis tool for individualized diagnosis and treatment of mental diseases.

[0006] The application adopts the following technical scheme:

[0007] The neural anatomical heterogeneity analysis method integrating the standard model and the non-negative matrix factorization comprises the following steps:

[0008] Step 1, preprocessing structural magnetic resonance image data, comprising: converting the collected structural magnetic resonance image data into a specified magnetic resonance imaging format (for example, NIfTI format), and performing segmentation processing on the structural image based on a magnetic resonance image processing tool (for example, FreeSurfer), dividing the brain into a plurality of brain regions based on an adopted brain region template (for example, Destrieux atlas), and extracting at least two brain morphological indexes of each brain region, and generating a plurality of n×1 column vectors based on each brain morphological index of all brain regions, wherein n represents the number of brain regions;

[0009] Preferably, the brain morphological indexes include cortical thickness and cortical surface area;

[0010] Step 2, respectively performing standard model migration on the brain morphological indexes of each brain region, using all healthy subject data of a center where the patient is located as an adaptation data set to construct a standard model of the center;

[0011] Inputting patient data into the migrated standard model to estimate the deviation score (i.e., Z score) of each patient in each brain morphological index of each brain region:

[0012] Wherein, the subscript parameter m is used to identify the patient, and the subscript parameter d is used to identify the brain region, that is, z md represents the deviation score of the mth patient in the brain region d, y md and are respectively the real and predicted brain morphological index values, is the noise variance caused by the estimated data uncertainty, is the variance caused by the modeling uncertainty;

[0013] Step 3, splicing the deviation scores z md of all brain regions of all patients to obtain the deviation score matrix Z d of each brain morphological index, and the matrix is n×M, wherein M represents the number of patients;

[0014] Splicing the deviation score matrices Z dThe components are concatenated along the n-direction to form a matrix Q with dimensions n×NM, where N represents the number of brain morphological indicators.

[0015] Extracting the non-positive parts from matrix Q and taking their absolute values ​​yields the cortical atrophy bias matrix Z (also known as the atrophy fraction matrix);

[0016] Step 4: Iterate the cortical atrophy bias matrix Z using orthogonal constraints and perform nonnegative matrix decomposition: Z = W × H + ∈ W, where W represents the brain region component matrix (size n × K, i.e., the number of brain regions n and the number of components K), H represents the weight matrix of each component on the subject (size K × Nm, i.e., the number of components K and N times the number of subjects 2m), and ∈ represents the residuals. The number of components K (also known as the number of factors) is a configurable variable, and its optimal value will be determined through subsequent stability analysis.

[0017] During the iteration process, the update formulas for the brain region component matrix W and the weight matrix H are set as follows:

[0018]

[0019] H = W T X

[0020] Here, ⊙ represents the Hadamard product, which is the element-wise multiplication of two matrices. This represents the element-wise division of two matrices. Both operations are performed at corresponding element positions, and the resulting matrix has the same dimensions as the input matrix.

[0021] When the preset iteration convergence conditions are met (such as convergence of iteration count, residual value, etc.), the iteration stops and proceeds to step 5;

[0022] Step 5: Calculate the semi-decomposition stability coefficient and reconstruction error of the non-negative matrix decomposition results to determine the optimal number of components K;

[0023] The column vectors of matrix Z are randomly shuffled and divided into N equal parts, namely Z′1~Z′. N Each part contains approximately the same number of subjects;

[0024] For each part Z′ r Perform nonnegative matrix decomposition based on orthogonal constraints to obtain the corresponding brain region component matrix W. r Where r = 1, ..., N, i.e., Z′ r =W r H r H r This is the weight matrix corresponding to the r-th part;

[0025] Calculate the component matrix W for each brain region r Component score matrix cWr wherein the score matrix cW r is the ith row jth column element of the score matrix cW is: are the ith column and jth column of the component score matrix cW r respectively;

[0026] The semi-decomposition stability coefficient SC is obtained based on the average of the correlation coefficients of the same row vectors between the component score matrices cW r ;

[0027] For the set K value range, the average of the semi-decomposition stability coefficients SC obtained by multiple acquisitions is taken as the final semi-decomposition stability coefficient SC of K avg , the higher the value, the more stable the decomposition result of the K value;

[0028] For each K value, the average of the multiple reconstruction errors is taken as the reconstruction error RE of K avg , the lower the reconstruction error, the better the decomposition effect of the K value;

[0029] The comprehensive measurement coefficient of each K value is obtained by integrating the SC avg and RE avg corresponding to each K value, and the K value corresponding to the optimal comprehensive measurement coefficient is the optimal component number K.

[0030] Preferably, the comprehensive measurement coefficient can be set as: αSC avg -βRE avg , wherein α and β are corresponding fusion coefficients.

[0031] When calculating the semi-decomposition stability coefficient SC, any two-by-two combination of the score matrices cW r corresponding to the N brain morphological indicators is performed, the correlation coefficients between the same row vectors of the two score matrices cW r under each combination are calculated, the row vectors of cW r represent different brain regions, and the correlation coefficient between the two cW r under the current combination is obtained based on the average of the n correlation coefficients, and the semi-decomposition stability coefficient SC is obtained by averaging the correlation coefficients of all combinations.

[0032] For example, for each K value, the SC value is repeatedly obtained num times, and the average of the SC values obtained each time is taken as the final semi-decomposition stability coefficient SC of the current K value.

[0033] Similarly, for the reconstruction error RE avg , the calculation formula can be represented as: wherein W * , H* The brain region component matrix and the weight matrix obtained when the iteration in step 4 is stopped are represented.

[0034] Further, in step 1, before the structural image is segmented, a motion correction process is further included to improve the accuracy of the obtained data; the segmentation of the structural image includes skull separation, gray matter and white matter segmentation, and cortical surface reconstruction.

[0035] Further, in step 2, the covariates used for the establishment and migration of the standard model include age, gender, and center. In order to improve the stability and robustness of the migrated model and reduce its sensitivity to data distribution assumptions, a hyperbolic sine inverse function transformation can be used to construct a warping likelihood function. Then, the transformation parameters of the pre-trained model are applied to the new data set to adjust the mean and variance in the latent Gaussian space, and the adjusted data is warped back to the original space. Finally, a warped linear Bayesian regression is used as the prediction method.

[0036] Further, in step 3, when the cortical atrophy bias matrix Z is obtained, each column of the bias score matrix Z d represents the bias score of a certain patient in all brain regions, and the same column number position corresponds to the same patient; each row represents the bias score of a certain brain region in all patients, and the same row number position corresponds to the same brain region; and all bias score matrices Z d are spliced into a matrix matrix Q along the number of patients M. That is, the obtained atrophy score matrix Z contains multiple morphological indicators such as cortical thickness and area, rather than a single indicator, which can more comprehensively represent the atrophy pattern of the patient's brain and capture the potential overlap and interdependence between indicators in subsequent analysis.

[0037] In step 4, the non-negative matrix factorization as an unsupervised multivariate analysis method can identify the spatial pattern of multiple morphological indicators and capture the pattern of cortical variability. Its output has the advantages of high interpretability, high sparsity, and high scalability across data sets.

[0038] In step 5, the optimal component number K obtained by the non-negative matrix factorization can determine the number of disease factors, and the disease factor represents the covariant pattern of the cortical morphological indicators of the disease, which can effectively explain and represent the cortical morphological atrophy pattern while preserving the individual heterogeneity of the subjects. Specifically, for a certain patient, the morphological atrophy pattern of the patient can be composed of one or more disease factors, and the proportion of different factors is represented by the weight.

[0039] Further, based on the optimal number of components K determined in step 5, the optimal brain region component matrix and weight matrix obtained by the orthogonal constraint based non-negative matrix factorization iteration on the cortical atrophy bias matrix Z can be used to construct detection features for the mental disease prediction task; and based on the detection features, the target mental disease prediction model is trained, thereby obtaining the target mental disease prediction model, which provides an auxiliary diagnosis tool for doctors.

[0040] The technical scheme provided by the present application at least brings the following beneficial effects:

[0041] The method provided by the present application is different from the prior art, and is aimed at the problem of individual heterogeneity of mental diseases. A neural dissection heterogeneity analysis method integrating a standard model and non-negative matrix factorization is proposed. Firstly, the standard model is established by model transfer, and the pre-training queue is trained on an independent large data set, thereby reducing the influence of abnormal values or sampling bias on the model. Secondly, the high-dimensional result of the standard model is decomposed by using the non-negative matrix factorization method, which can integrate and capture the covariant mode of different morphological indicators, reveal their interdependent relationship, and output disease factor results with high interpretability, high sparsity and high cross-data set scalability, thereby facilitating further analysis and interpretation. BRIEF DESCRIPTION OF DRAWINGS

[0042] The above and / or additional aspects and advantages of the present application will become apparent and more readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings, in which:

[0043] Figure 1 A step flowchart of a neural dissection heterogeneity analysis method integrating a standard model and non-negative matrix factorization proposed for the embodiments of the present application.

[0044] Figure 2 An example output of the result of the conventional model.

[0045] Figure 3 A conventional model brain map of an example patient.

[0046] Figure 4 A bias score matrix of the patient after splicing.

[0047] Figure 5 An atrophy score matrix of the patient extracted.

[0048] Figure 6 Variation curves of the average half-decomposition stability coefficient SC and the average reconstruction error RE in the optimal K selection process. avg avg

[0049] Figure 7 ​​Decompose the 8 disease factors of the cortical atrophy bias matrix Z of the patient in the specific implementation method.

[0050] Figure 8 The result corresponding to the weight composition. DETAILED DESCRIPTION

[0051] In order to make the person skilled in the art better understand the technical solutions in the specification, the technical solutions of the embodiments of the present application will be described in detail and completely below in combination with the drawings in the embodiments of the present application. Obviously, the embodiments described by referring to the drawings are exemplary and are intended to explain the present application, and cannot be understood as a limitation of the present application.

[0052] The embodiment of the present application provides a neural anatomy heterogeneity analysis method integrating standard models and non-negative matrix factorization, which is a brain structure atrophy pattern analysis method based on combination of a standard model and an orthogonal constrained non-negative matrix factorization, so as to better study the spatial distribution pattern of the mental disease population level while retaining individual heterogeneity. The method provided in the present application obtains the deviation scores of the cortical thickness and surface area of each brain region of the patient by using a pre-trained standard model. Then, the negative part of the deviation score is extracted and spliced to form a cortical atrophy pattern matrix. Finally, the orthogonal constrained non-negative matrix factorization is used to reduce the dimension of the matrix to reveal the atrophy pattern of the cerebral cortex structure and effectively extract the spatial distribution pattern at the population level while retaining the individual heterogeneity information. The method can accurately identify the abnormal variation of the brain structure, provide a new data analysis tool for individualized diagnosis and treatment of mental diseases, and has high interpretability and scalability across datasets.

[0053] Reference Figure 1 In one embodiment, the neural anatomy heterogeneity analysis method integrating standard models and non-negative matrix factorization provided by the present application comprises the following steps:

[0054] Step 1: Preprocessing of structural magnetic resonance image

[0055] The data set of the embodiment includes data of 224 healthy subjects from 6 centers and data of 684 Alzheimer's disease (AD) patients from corresponding 4 centers. The structural magnetic resonance data is segmented by Freesufer (Version 7.2) software, Destrieux atlas is used as the whole brain cortex template, the whole brain is divided into 148 brain regions, the estimated values of cortical thickness and surface area (aparc.a2009s.stat) are extracted from the output folder of each subject, the cortical thickness and surface area indicators of each brain region are calculated, and two 148x1 column vectors are obtained for each subject.

[0056] Step 2: Transfer of the standard model

[0057] First, using the pcntoolkit toolkit, a pre-trained standard model from approximately 58,000 healthy individuals across 82 sites was transferred to AD patient data. Covariates selected during the transfer included age, sex, and center. A standard model for each center was constructed using data from all healthy participants across four centers containing patient data as the adaptation dataset. Patient data was then input into the transferred standard model to estimate the cortical thickness bias score and cortical surface area bias score (i.e., Z-score) for each brain region in 684 patients, as shown in the following formula:

[0058]

[0059] Among them, z md y represents the deviation score of the m-th subject in brain region d. md and These are the actual and predicted brain morphology index values, respectively. This is the noise variance caused by the estimated data uncertainty. To model the variance caused by uncertainty. The standard model outputs the cortical thickness bias score for one brain region as follows: Figure 2 As shown, fractional brain maps of cortical thickness deviation in some patients are as follows: Figure 3 As shown, the main atrophy areas in most patients are concentrated in the medial and lateral temporal lobes.

[0060] Step 3: Construct the shrinkage bias matrix

[0061] The deviation scores are concatenated into a cortical thickness deviation score matrix Z1 and a cortical surface area deviation score matrix Z2, both with a size of 148×684. These two matrices are then horizontally concatenated into the form [Z1, Z2]. Figure 4 As shown, the non-positive parts are separated to obtain a cortical atrophy bias matrix Z of size 148×1368, as shown. Figure 5 As shown, this serves as the input for subsequent steps.

[0062] Step 4: Nonnegative matrix decomposition based on orthogonality constraints

[0063] The cortical atrophy bias matrix Z generated in step 3 is subjected to nonnegative matrix decomposition based on orthogonal constraints, as shown in the following formula.

[0064] Z=W×H+∈W (2)

[0065] Where W represents the brain region component matrix (size 148×K), H represents the weight matrix of each component on the subject (size K×1368), and ∈ represents the residuals. The number of factors K is a configurable variable, and its optimal value will be determined through subsequent stability analysis.

[0066] In the iteration process, the formula for updating W and H is as follows

[0067]

[0068] H = W T X

[0069] where denotes the Hadamard product, i.e., the element-wise multiplication of two matrices, and denotes the element-wise division of two matrices. Both operations are performed at the corresponding element positions, and the resulting matrix has the same dimension as the input matrix.

[0070] Using the above method, the cortical atrophy bias matrix Z is decomposed, and the atrophy bias pattern of AD is divided into component factor matrix W (148xK) and its corresponding weight matrix H (Kx1368). The matrix H is divided into left cortical thickness (Kx684) and right cortical surface area (Kx684).

[0071] Step 5: Calculate the average semi-decomposition stability coefficient SC avg

[0072] First, the column vectors of the cortical atrophy bias matrix Z are randomly shuffled and equally divided into Z1 and Z2, each containing 342 subjects. Orthogonal constraint-based non-negative matrix decomposition is performed on the two parts to obtain component factor matrices W1 and W2.

[0073]

[0074] Then, the cosine similarity matrices cW1 and cW2 (148x148) between W1 and W2 are calculated, which represent the similarity between brain regions in the component factors.

[0075]

[0076] where cw ij is the i-row j-column element of cW, w i and w j are the i-th column and j-th column of W, respectively.

[0077] Finally, the average of the correlation coefficients coef between the corresponding rows of cW1 and cW2 is calculated as the semi-decomposition stability coefficient SC to evaluate the stability of the component number K as the number of brain regions.

[0078]

[0079] For K ranging from 2 to 30, the above work is repeated 100 times, and the average SC of SC is calculated.avg As the final semi-decomposition stability coefficient of K, the higher the stability coefficient, the more stable the decomposition result of K value, and the final SC avg value curve with K value is shown in Figure 5 .

[0080]

[0081] Step 6: Calculate the average reconstruction error RE avg

[0082] Calculate the reconstruction error RE of each non-negative matrix decomposition, which is defined as the two-norm of the difference between the original matrix Z and WH

[0083] RE = ||Z-WH|| (8)

[0084] For each K from 2 to 30, the same operation is repeated 100 times in this embodiment, and the average value RE of the decomposition error is calculated avg As the reconstruction error of K value, the lower the reconstruction error, the better the decomposition effect of K value, and the final SC avg value curve with K value is shown in Figure 5 .

[0085]

[0086] Step 7: Determination of optimal component number K

[0087] By comprehensively considering the average semi-decomposition stability coefficient SC avg and the average reconstruction error RE avg , the optimal K value is determined to be 8 as the component number of non-negative matrix decomposition, and the cortical atrophy deviation matrix Z of AD is decomposed into 8 brain region components as shown in Figure 6 , in which factor 1 is mainly the parietal lobe part; factor 2 is mainly the frontal lobe part; factor 3 is mainly the occipital lobe part; factor 4 is mainly some scattered areas including the anterior cingulate gyrus structure; factor 5 is mainly the medial temporal lobe part including the parahippocampal gyrus structure; factors 6 and 7 are mainly the left and right lateral temporal lobe parts respectively, including the superior temporal sulcus, angular gyrus, middle temporal gyrus, etc. structure; factor 8 is mainly some scattered areas including the posterior cingulate gyrus, distance sulcus, etc. structure. The corresponding weight composition is shown in Figure 7 , and the weight composition of each patient individual can be considered as the composition ratio and content of disease factors, which reflects the individual heterogeneity of patient brain morphological atrophy pattern.

[0088] In the description of the specification, the description of the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. In addition, those skilled in the art can combine and combine the different embodiments or examples described in the specification and the features of the different embodiments or examples without contradiction, and the scope of the preferred embodiments of the present application includes additional implementations.

[0089] Any process or method descriptions or descriptions of the flow diagrams in the specification can be understood as representing code modules, segments, or portions of code which include one or more executable instructions for implementing the specified logic functions (or steps) of the process, and the preferred embodiments of the present application include additional implementations in which the order of execution of the executable instructions is changed, additional or fewer processes are performed, or the functions are performed in different sequences, as will be appreciated by those skilled in the art.

[0090] It should be understood that parts of the present application can be implemented in hardware, software, firmware, or a combination thereof. In the above-described embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. As in another embodiment, if implemented in hardware, any of the following technologies or their combinations can be used: discrete logic circuit with logic gates for implementing logic functions on data signals, application specific integrated circuit with suitable combination logic gates, programmable gate array (PGA), field programmable gate array (FPGA), etc.

[0091] Those skilled in the art of the present technology can understand that all or part of the steps carried out by the above-described embodiment method can be completed by a program instructing the relevant hardware, and the program can be stored in a computer readable storage medium. The program, when executed, includes one or a combination of steps of the method embodiment.

[0092] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method of neuroanatomical heterogeneity analysis integrating standard models and non-negative matrix factorization, characterized in that, The method comprises the following steps: Step 1, preprocessing structural magnetic resonance imaging data, including: converting the collected structural magnetic resonance imaging data into a specified magnetic resonance imaging format, and segmenting the structural image based on a magnetic resonance image processing tool, dividing the brain into multiple brain regions based on the brain region template used, and extracting at least two brain morphology indicators for each brain region, and generating multiple n x 1 column vectors based on each brain morphology indicator of all brain regions, wherein n represents the number of brain regions; Step 2, migrating the standard model for each brain morphology indicator of each brain region, using all healthy subject data of the center where the patient is located as an adaptive data set to construct the standard model of the center; The patient data is inputted into the migrated standard model to estimate the bias score of each patient for each brain morphological indicator in each brain region: wherein the subscript parameter m is used to identify the patient, the subscript parameter d is used to identify the brain region, z md represents the bias score of the mth patient in the brain region d, y md and are the true and predicted brain morphological indicator values, respectively, is the noise variance caused by the estimated data uncertainty, is the variance caused by the modeling uncertainty; Step 3, concatenate the deviation scores z for all brain regions for all patients md obtain the deviation score matrix Z for each brain morphological index d , which has dimension n x M, where M represents the number of patients; The deviation score matrix Z of all brain morphology indexes is calculated as follows: d The n-dimensional matrix Q with the dimension of n x NM is spliced along the n direction, wherein N represents the number of brain morphology indexes. Extracting the non-positive part from the matrix Q and taking the absolute value to obtain the cortical atrophy deviation matrix Z; Step 4: iteratively performing orthogonal constraint-based non-negative matrix decomposition on the cortical atrophy deviation matrix Z: Z = W x H + ∈W, wherein W represents a brain region component matrix with a dimension of n x K, K represents the number of components, H represents a weight matrix of each component on the subjects with a dimension of K x Nm, and ∈ is a residual error; In the iteration process, the update formula of the brain region component matrix W and the weight matrix H is set as: H = W T X where denotes the Hadamard product, denotes element-wise division of two matrices, and the superscript T denotes transposition. When the preset iteration convergence condition is met, the iteration is stopped and step 5 is entered; Step 5: calculating the semi-decomposition stability coefficient and the reconstruction error of the result of the non-negative matrix decomposition to determine the optimal component number K. The column vectors of matrix Z are randomly shuffled and equally divided into N parts, i.e. Z'1 ~ Z'N N each of which contains approximately equal number of subjects; perform non-negative matrix factorization based on orthogonal constraints on each part Z' r to obtain the corresponding brain region component matrix W r where r = 1, …, N, i.e. Z' r = W r H r , H r is the weight matrix corresponding to the rth part. computing a component score matrix cW r for each brain region component matrix W r where the score matrix cW r is computed as are the i-th column and j-th column of the brain region component matrix W r respectively.​ Based on the composition score matrix cW r The mean of the correlation coefficients of the same row vectors between the matrices cW and cW+1 gives the semi-decomposed stability coefficient SC; For the set range of K values, the mean value of the semi-decomposed stability coefficients SC obtained from multiple acquisitions is taken as the final semi-decomposed stability coefficient SC of K avg ; For each value of K, the mean of the multiple reconstruction errors is taken as the reconstruction error RE for K avg ; Summarize the SC corresponding to each K value avg and RE avg Get the summary metric coefficient of each K value, and the K value corresponding to the optimal summary metric coefficient is the best component number K.

2. The method of claim 1, wherein, The brain morphology indicators include cortical thickness and cortical surface area.

3. The method of claim 1, wherein, The comprehensive metric coefficient is set as: aSC avg - βRE avg Wherein, a, β are corresponding fusion coefficients.

4. The method of claim 1, wherein, reconstruction error where W * , H * denote the brain region component matrix and the weight matrix obtained when the iteration is stopped in step 4, and num denotes the number of repetitions.

5. The method of claim 1, wherein, In step 1, before segmenting the structural image, motion correction processing is also included; segmenting the structural image includes skull separation, gray matter and white matter segmentation, and cortical surface reconstruction.

6. The method of claim 1, wherein, In step 2, the covariates used for establishing and migrating the standard model include age, gender, and center.

7. The method of claim 1, wherein, In step 3, the deviation score matrix Z is obtained by cortical atrophy deviation matrix Z d Each column of the deviation score matrix Z represents the deviation score of a certain patient in all brain regions, and the positions with the same column number correspond to the same patient; each row represents the deviation score of a certain brain region in all patients, and the positions with the same row number correspond to the same brain region; and all deviation score matrices Z d are spliced into a matrix matrix Q along the number of patients M direction.

Images