Method and apparatus for evaluating a variational dependence

Abstract

The application relates to a variational correlation evaluation method and device, which comprises the following steps: acquiring magnetic resonance image data to obtain to-be-processed data; determining an iteration position sequence of the to-be-processed data and extracting iteration features to obtain a training set and a test set; inputting the training set into a variational correlation evaluation classification model for iteration training and testing, judging whether the trained variational correlation evaluation classification model converges or not, and obtaining the variational correlation evaluation classification model when the model converges, and performing effective feature extraction based on the variational correlation evaluation classification model; determining each effective feature position sequence, converting the effective feature position sequence into a three-dimensional brain structure matrix, covering the three-dimensional brain structure matrix to a preset standard human brain template, and identifying effective features related to each stimulation condition. The application quantifies the contribution of a single voxel in the process of executing a specific cognitive function, and finally identifies and extracts the least amount of features that can best represent the target stimulation condition.

Description

Technical Field

[0001] This invention belongs to the field of big data technology, specifically relating to a variational correlation assessment method and apparatus. Background Technology

[0002] In the field of cognitive neuroimaging, traditional research on cognitive mechanisms has focused on extracting topologically activated brain regions or networks under specific conditions. However, the working mechanism of the brain is not a one-step process, but a cognitive process composed of cognitive states at multiple consecutive time points.

[0003] In related technologies, common "representation" research methods include univariate analysis and multivariate pattern analysis. Univariate analysis assumes that voxels act independently, constructing a linear relationship between BOLD (Blood Oxygen Level-Dependent) signals and stimulus conditions. Univariate analysis is relatively simple to compute and requires fewer computational resources, making it suitable for large-scale spatial analysis, especially for higher cognitive functions that require comprehensive consideration of the entire brain's neural activity. However, this approach of ignoring the connections between different voxels and independently considering the relationship between a single voxel and a stimulus has limitations: this approach contradicts the physiological context of the brain functioning as a whole; ignoring the spatial relationships between different cortical regions may lead to information loss. In particular, the spatial averaging method involved in univariate analysis removes voxels that respond weakly to specific conditions. These voxels may carry characteristics indicating whether the brain is processing or not processing certain stimulus information, thus blurring the fine spatial boundaries useful for distinguishing the brain's functional topological regions corresponding to different stimulus conditions. Unlike univariate analysis, multivariate pattern analysis effectively utilizes the correlations between candidate voxel features to uncover hidden features beneath higher-resolution imaging signals with greater precision than the original data's spatial resolution. For data acquired through non-invasive methods, the non-isotropic arrangement of brain structures and blood flow throughout the brain makes this level of feature selection difficult, if not impossible, with traditional data analysis methods. Another difference from univariate analysis is that if the differences in cognitive states behind each whole-brain neuroimaging data point are sufficiently large, multivariate pattern analysis can analyze the cognitive states behind individual neuroimaging signals. In other words, multivariate pattern analysis can analyze cognitive states during experimental stimulus processing with high temporal resolution. However, considering the potential for overfitting due to the high dimensionality and small dataset characteristics of functional magnetic resonance imaging (fMRI) data, regardless of the chosen algorithm for classification or regression models, the primary challenge is data dimensionality reduction. Especially for machine learning algorithms with inherently complex structures, the limited availability of neuroimaging data necessitates a finite number of features for model training; in this case, a more complex model structure increases the risk of overfitting.

[0004] In addition to analyzing brain "representations," a complete study of cognitive processes cannot be separated from exploring brain "operations." For the problem of information "operations," the most widely used approach is the brain network method. Compared to the functional labeling of brain regions in "representation" methods, the brain network method focuses on functional labeling of information exchange between multiple brain regions. This functional connectivity approach analyzes the functional correlation between different brain regions from a macroscopic perspective, explaining different cognitive processes from different brain connectivity patterns. Resting-state networks (RSNs) are defined as regions that have a functional tendency to exchange information in a resting state; task-state networks directly link network topology and cognitive function through computational logic similar to resting-state networks. However, neither can quantify the contribution of each region or the cross-regional information and control flow during task execution. Defined brain regions may simultaneously play different roles under specific cognitive functions. The analytical characteristics of brain network methods for information interaction across multiple brain regions make it impossible to segment the different functional response "representations" of a single brain region at a microscopic scale. Furthermore, brain network methods are limited to quantifying the connection strength between brain regions within a network, and cannot effectively assess the state changes of each brain region itself. Considering individual differences, it is unclear whether the dynamic changes of each brain region differ among different individuals under specific external stimuli.

[0005] In summary, traditional univariate analysis methods focus on the "representational" localization of activated brain regions, neglecting the information interaction between voxel features, resulting in static activation locations distributed across all levels of brain structure. Brain network methods, on the other hand, sacrifice high-resolution functional analysis of individual features, enabling the description of the interactive "operational" processes between different parts of the brain. Therefore, traditional analytical methods can only make trade-offs between "representational" location localization and "operational" functional connectivity, unable to simultaneously leverage the advantages of both to quantitatively explain the relationship between neuroimaging signals and specific stimulus conditions. Thus, understanding how "operation" and "representation" dynamically interact during cognition is a pressing issue that needs to be addressed. Summary of the Invention

[0006] In view of this, the purpose of the present invention is to overcome the shortcomings of the prior art and provide a variational correlation assessment method and apparatus to solve the problem of how "operation" and "representation" interact dynamically in the cognitive process.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: a variational correlation assessment method, comprising:

[0008] Magnetic resonance imaging data is acquired and preprocessed to obtain data to be processed; wherein, the magnetic resonance imaging data includes brain voxel features;

[0009] Determine the iteration position sequence of the data to be processed, extract the iteration features of the data to be processed, divide the iteration features according to a preset ratio, and obtain a training set and a test set;

[0010] The training set is input into the variational correlation assessment classification model for iterative training. The variational correlation assessment classification model is tested after reaching the training number to determine whether the trained variational correlation assessment classification model has converged. Effective feature extraction is performed when it converges.

[0011] The effective feature location sequence related to each stimulus is determined, the effective feature location sequence is converted into a three-dimensional brain structure matrix, and the three-dimensional brain structure matrix is ​​overlaid on a preset standard human brain template to identify the effective features related to each stimulus condition.

[0012] Furthermore, effective feature extraction is performed based on the variational correlation assessment classification model, including:

[0013] Determine the effective feature sparse position sequence for this iteration, and determine the sub-position sequence for calculating the features in the next iteration;

[0014] Determine whether the effective feature sparse position sequence of this iteration is the same as the sub-position sequence of the feature for the next iteration.

[0015] If they are different, proceed to the next calculation iteration; otherwise, abandon the current iteration, increase the number of feature data to be used in the calculation, and restart the calculation iteration.

[0016] Furthermore, the i-th effective feature sparse location sequence is composed of the sub-location sequence ζ used to locate the effective feature locations in the local dimension. i Composed of supplementary sequences;

[0017] The supplementary sequence is used to convert the local dimension of the sub-position sequence into the global dimension of the position sequence.

[0018] Furthermore, the variational correlation assessment classification model includes: a mean layer and a variance layer set in parallel;

[0019] The average layer is used to evaluate the correlation between weight values ​​and specific experimental conditions, taking into account the interaction effects of multiple variables.

[0020] The variance layer is used to evaluate the robustness of the averaging layer.

[0021] Furthermore, the variational correlation assessment classification model includes:

[0022]

[0023]

[0024] Among them, X Train For the training input data, y Train Let w be the predicted class label corresponding to the training data, and b be the weight matrix and bias matrix, respectively. Q(w|X) Train ,y Train Let P(w) be the posterior distribution function and P(w) be the joint prior distribution function. These are the prior average parameters. Let be the prior variance parameter, ⊙ be the Hadamard product, and D be the... KL [Q(w|X Train ,y Train )||P(w)] is the KL divergence.

[0025] Furthermore, determining whether the trained variational correlation assessment classification model has converged includes:

[0026] Determine whether the number of training iterations in a single iteration is greater than or equal to the preset number of iterations;

[0027] If so, perform a model convergence test. If the classification accuracy of the test set is higher than the preset accuracy, the model is determined to be converged; otherwise, the model is determined not to be converged.

[0028] If the model fails to converge, iterative training continues. If the model still fails to converge after a preset number of training iterations, the current iteration ends, and the number of features required for the next iteration is increased.

[0029] Furthermore, it also includes:

[0030] The target category is input into the variational correlation assessment classification model to obtain the prediction vector;

[0031] The sub-correlation coefficient is calculated based on the prediction vector; wherein the sub-correlation coefficient is the percentage of the difference between each voxel and the linear result calculated with the weight values ​​corresponding to different categories in the two parallel regression layers in the difference between the first element and any other element of the prediction vector.

[0032] The cognitive state matrix is ​​obtained based on the sub-correlation coefficients; the cognitive state matrix is ​​used to represent the time-domain changes of whole-brain voxels.

[0033] Furthermore, the preprocessing includes pre-preprocessing and post-preprocessing;

[0034] The preprocessing includes inter-layer temporal correction, head motion correction, image spatial registration, and spatial normalization.

[0035] The post-preprocessing includes noise reduction and normalization.

[0036] Furthermore, a batch feature input method is adopted to input the whole brain voxel features participating in the calculation iteration in batches and participate in the feature extraction calculation iteration sequentially;

[0037] The iterative position sequence is a sparse sequence.

[0038] This application provides a variational correlation assessment apparatus, including:

[0039] The acquisition module is used to acquire magnetic resonance imaging data and perform preprocessing to obtain data to be processed; wherein, the magnetic resonance imaging data includes brain voxel features;

[0040] An extraction module is used to determine the iteration position sequence of the data to be processed, extract the iteration features of the data to be processed, divide the iteration features according to a preset ratio, and obtain a training set and a test set.

[0041] The training module is used to input the training set into the variational correlation assessment classification model for iterative training, and to test the variational correlation assessment classification model after reaching the training number, to determine whether the trained variational correlation assessment classification model has converged, and to extract effective features when it converges.

[0042] The identification module is used to determine the effective feature location sequence related to each stimulus, convert the effective feature location sequence into a three-dimensional brain structure matrix, and overlay the three-dimensional brain structure matrix onto a preset standard human brain template to identify the effective features related to each stimulus condition.

[0043] The beneficial effects that can be achieved by adopting the above technical solution in this invention include:

[0044] This invention provides a variational correlation assessment method and apparatus. The method provided in this application quantifies the contribution of a single voxel in the execution of a specific cognitive function. Compared with traditional methods, it eliminates the need for feature pre-selection before model training, allowing the whole brain to be treated as a whole and simultaneously analyzing the features of all brain voxels. Furthermore, through a batch feature input algorithm, a balance is cleverly found between ensuring computational efficiency and avoiding overfitting. Two sets of adversarial strategies, implemented by two parallel linear regression layers, ultimately achieve the synchronous execution of feature selection and model training. Each feature selection iteration requires determining the features participating in the current iteration based on the effective feature position sequence obtained from the previous iteration, ultimately identifying and extracting the feature with the smallest volume that best represents the target stimulus condition. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a schematic diagram illustrating the steps of the variational correlation assessment method of the present invention;

[0047] Figure 2 This is a flowchart illustrating the variational correlation assessment method of the present invention.

[0048] Figure 3 This is a schematic diagram illustrating the training and testing of the variational correlation assessment classification model provided by the present invention.

[0049] Figure 4 (a) is the loss equation provided by this invention. Figure 4 (b) is a schematic diagram of the distribution of the weight values ​​of the probability density;

[0050] Figure 5 This is a schematic diagram illustrating the effective feature extraction provided by the present invention;

[0051] Figure 6 A schematic diagram illustrating the effective feature recognition related to stimulus conditions provided by this invention;

[0052] Figure 7 (a) is a schematic diagram of the correlation coefficient provided by the present invention. Figure 7 (b) is a schematic diagram of the cognitive state matrix; Figure 7 (c) is a dynamic diagram of cognitive states;

[0053] Figure 8 This is a schematic diagram of a verification experiment for the variational correlation assessment method of the present invention;

[0054] Figure 9 This is a schematic diagram of the variational correlation evaluation device of the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0056] The following describes a specific variational correlation evaluation method and apparatus provided in an embodiment of this application, with reference to the accompanying drawings.

[0057] like Figure 1 As shown, the variational correlation assessment method provided in this application embodiment includes:

[0058] S101, acquire magnetic resonance imaging data and perform preprocessing to obtain data to be processed; wherein, the magnetic resonance imaging data includes brain voxel features;

[0059] In some embodiments, the preprocessing includes pre-preprocessing and post-preprocessing;

[0060] The preprocessing includes inter-layer temporal correction, head motion correction, image spatial registration, and spatial normalization.

[0061] The post-preprocessing includes noise reduction and normalization.

[0062] Among them, image layer time correction requires that the acquisition time point of a specific data layer be the same within different TR acquisition cycles; head movement correction assesses the impact of head movement on brain region localization; spatial registration registers functional phase data to structural phase data; and standardization fits the registered data to a standard T1 template through a nonlinear equation.

[0063] Slice Timing: Functional magnetic resonance imaging (fMRI) scanning methods are divided into slice-by-slice scanning and slice-by-slice scanning. In event-related experimental paradigms, slice-by-slice scanning is usually chosen. With this method, the acquired MRI images generally undergo time-slice correction before head motion correction. If slice-by-slice scanning is used, the corresponding fMRI data undergoes head motion correction before time correction. Regardless of the scanning method, it is essential to ensure that the images describing the whole-brain BOLD response in different sampling periods are acquired at the same time points. The BOLD response signals acquired in several slices within each sampling period actually correspond to the neural response activity of each brain slice acquired at a number of time points equal to the number of scanned slices. In practice, this acquisition method exhibits slice-by-slice acquisition time drift, which necessitates slice-by-slice timing correction to correct the differences in acquisition time points between different whole-brain volumes. Under the premise of relatively short sampling periods (TR≤2s) and slice-by-slice data acquisition, timing correction can ensure good robustness.

[0064] Head movement correction (Realignment): Head movements are inevitable during experiments conducted in a functional magnetic resonance imaging (fMRI) environment. These movements are independent of the experimental paradigm and depend on the subject's physiological state during the experiment; for example, swallowing can lead to significant head movements. Head movement correction is a crucial step in the preprocessing of MRI image data. If head movements exceed a certain threshold, they can cause biases in the localization of activated brain regions in fMRI images. Without assessing head movements, some voxels within the final extracted activation feature clusters that do not belong to specific brain regions may be located in those regions, resulting in incorrect functional activation location localization and subsequent errors in mechanism analysis. This error primarily occurs at the edges of functional phase data. The purpose of head movement correction is to assess whether head movements cause misalignment in fMRI imaging.

[0065] Head motion correction is based on rigid transformation. It assumes the subject's head is a rigid body, meaning it will not undergo translation or rotation in any of the three directions (i.e., all six degrees of freedom are zero). Head motion correction requires selecting one reference image from multiple time-series images within a TR sampling period. Other time-series images are then matched against the reference image layer to calculate translation and rotation parameters, thus achieving head motion correction for all images within the same sampling period. Typically, the first layer or an intermediate image layer from the acquisition time-domain sequence is chosen as the reference image.

[0066] After the head motion correction calculation is completed, the head motion parameters are evaluated. If the head motion parameters are within the allowable threshold range, the sample data is retained; otherwise, the data is deleted. The threshold for the head motion parameters is adjusted appropriately according to the sample size. When the sample size is large, the threshold standard is increased to ensure the accuracy of the experimental results. Generally, if the translation on the X, Y, and Z axes is less than 3 mm, and the deflection angle on each axis is less than 1°, that is, the error is less than the size of one voxel, it is considered that the head motion has no impact on the data, and subsequent analysis can be carried out.

[0067] Image spatial registration (Coregister): The main task is to superimpose low-resolution echoplanar images (EPI) onto high-resolution anatomical images. Due to magnetic susceptibility and artifact effects causing geometric distortions and activation intensity changes in functional images, deconvolution correction is required to improve registration accuracy. Image spatial registration employs a Bayesian maximum a posteriori (MAP) estimation algorithm, combined with head morphology and size, to achieve accurate registration between various types of functional and anatomical data using intermediate images. Image spatial registration aligns every voxel in all images, completing the information transformation between functional and structural data. For the same subject, the functional and structural phases have a linearly related translational and rotational relationship, without distortion. After registration, a matrix appears in the structural image's HDR file, containing the information exchange between the functional and structural phases.

[0068] Spatial Normalization: A key aspect of functional magnetic resonance imaging (fMRI) visualization is accurately mapping the detected functional phase activation regions onto high-spatial-resolution structural phases. Since functional phase activation maps lack anatomical information, they cannot be directly registered to anatomical structural phases. However, after registration, the functional phase maps and structural phases can share a coordinate system. Specifically, this involves using the transformation matrix between the registered functional phase maps and the anatomical structural phases. The convolution matrix between the average image generated by registration or the registered anatomical structural phase and the standard anatomical spatial template image is calculated, and this convolution matrix is ​​used to perform coordinate transformation for each functional phase data point. This ensures that functional phase data from different samples or modalities are studied within the same coordinate system.

[0069] In specific calculations, considering individual differences among subjects, when different brain structures are processed uniformly using the same coordinate transformation method, the rigid transformation in the registration stage is no longer applicable. In this case, it is necessary to transform the brain spaces of different subjects onto a unified standard human brain template through affine transformation of overall deformation and local nonlinear transformation. As mentioned earlier, the MNI template defined by the Montreal Neurological Institute is usually chosen as the standard template in the field of brain topology research; of course, other standard templates are available besides the MNI template. The spatial normalization algorithm aims to find a spatial transformation matrix that minimizes the Euclidean distance between the target functional phase image and the reference template image. Under the premise of satisfying this objective, it strives to ensure that the contrast of the target functional phase image is as similar as possible to the contrast of the reference template, and completes the unbiased estimation of the spatial transformation parameters to obtain the optimal solution.

[0070] Post-preprocessing: Following the pre-preprocessing of SPM, the raw EPI data was converted into whole-brain functional magnetic resonance imaging (fMRI) data comprising 592,895 voxels (each whole-brain volume contains 79 × 95 × 79 voxels). After SPM pre-preprocessing, post-preprocessing began. First, based on the time series of external stimuli throughout the experiment, all fMRI data acquired during rest intervals were averaged to serve as the baseline signal for the corresponding task execution run. Then, the baseline signal was subtracted from each whole-brain data acquired during the stimulation phase of the run to minimize the influence of environmental noise. Next, all voxels in each whole-brain volume were normalized to conform to a standard Gaussian distribution. Finally, all normalized whole-brain data matrices were reshaped into matrices with the new data structure, and then all task-state whole-brain data within a task execution run were stacked into a matrix with M rows (M whole-brain data points) and N columns (each brain containing N voxel features from that acquisition).

[0071] The goal of preprocessing is to verify the feasibility of the original data and improve resolution; the goal of postprocessing is to eliminate noise interference and achieve preliminary feature normalization so that the subsequent model calculation stage can converge as quickly as possible. Finally, each whole-brain functional magnetic resonance imaging (fMRI) data matrix is ​​reshaped into an array. This is because the variational correlation assessment algorithm architecture only studies the mathematical correlation between voxels or cortical regions (voxel clusters), rather than the different spatial relationships between cortical regions and the correlation before and after the acquisition time point. In other words, it is assumed that any cognitive state can be converted into an N-dimensional vector from the three-dimensional whole-brain fMRI data acquired at each corresponding data acquisition point. Therefore, the i-th row of the final matrix represents all N features of the corresponding i-th whole-brain fMRI data, while the j-th column represents an M-dimensional vector composed of the j-th voxel features from M different fMRI whole-brain images.

[0072] S102, determine the iteration position sequence of the data to be processed, extract the iteration features of the data to be processed, divide the iteration features according to a preset ratio, and obtain the training set and the test set;

[0073] In this application, the model needs to be trained at least 1000 times during the model training and testing phases of each computational iteration.

[0074] A subposition sequence is a vector representing the feature positions evaluated during a completed computational iteration. The subposition sequence is also a sparse sequence, where positions equal to "1" contain positions evaluated as valid features in the completed iterations and positions of new features needed to compensate for max_fea in order to improve computational efficiency; these features will be used in the next computational iteration. Conversely, elements in the subposition sequence that are "0" represent positions corresponding to voxels deemed meaningless in the completed computational iterations.

[0075] That is, the position sequence of the i-th computation iteration is composed of the sub-position sequence ζ used to locate the effective feature positions in the local dimension. i (Originally obtained from the (i-1)th iteration, when i = 1, ζ1 is an array of shape [1×n], where n = max_fea and all its elements are equal to 1) and a supplementary sequence (where all elements are 0, representing the positional elements corresponding to the remaining features that have not yet participated in the iteration) This indicates a concatenation operation. It is precisely due to the supplementary sequence that the local dimension of the sub-positional sequence is transformed into the global dimension of the positional sequence. In the i-th data selection phase, the required positional sequence is derived from the sub-positional sequence ζ. i (The number of elements at the position of "1" is equal to max_fea by default, and the sequence shape is [1*n]) and the supplementary sequence (all elements are at position "0", and the sequence shape is [1*(Nn)]).

[0076] After the location sequences are constructed, the next step is to divide the data matrix obtained in the data preprocessing stage into two non-overlapping data matrices according to a specific ratio (80% / 20% in this study on variational correlation assessment of cognitive processes). The training set is used for model training, and the test set is used for model generality testing. The batch features required in the next computation iteration are the Hadamard products of these two data matrices and the aforementioned location sequences (where ] represents the Hadamard product). The features required for the next computation iteration consist of the effective features obtained from the completed iterations and new, uncomputed features added to compensate for max_fea.

[0077] S103, the training set is input into the variational correlation assessment classification model for iterative training, and the trained variational correlation assessment classification model is tested to determine whether the trained variational correlation assessment classification model has converged, and effective feature extraction is performed when it converges.

[0078] In some embodiments, a batch feature input method is used to input the whole brain voxel features participating in the calculation iteration in batches and participate in the feature extraction calculation iteration sequentially;

[0079] The iterative position sequence is a sparse sequence.

[0080] Specifically, in order to strictly control the number of features involved in each calculation iteration, a batch feature input algorithm was developed based on the variational Bayesian framework. This algorithm inputs whole-brain voxel features participating in the calculation iteration in batches and sequentially participates in the feature extraction calculation iteration, so as to select the features most significantly correlated with the stimulus conditions by selecting whole-brain voxels in multiple batches.

[0081] In some embodiments, effective feature extraction is performed based on the variational correlation assessment classification model, including:

[0082] Determine the effective feature sparse position sequence for this iteration, and determine the sub-position sequence for calculating the features in the next iteration;

[0083] Determine whether the effective feature sparse position sequence of this iteration is the same as the sub-position sequence of the feature for the next iteration.

[0084] If they are different, proceed to the next calculation iteration; otherwise, abandon the current iteration, increase the number of feature data to be used in the calculation, and restart the calculation iteration.

[0085] Understandably, this application sets a maximum number of features (max_fea) that need to be computed in parallel for each feature extraction iteration. If the number of voxels that were not previously involved in feature validity evaluation is sufficient to compensate for the maximum number of features, the number of features required for each feature extraction iteration is equal to max_fea.

[0086] Understandably, during model training, n-dimensional features are determined in the data matrix to participate in the i-th model training and testing. For example... Figure 3 As shown, X Train Y represents the training data matrix in the i-th feature extraction iteration, with dimensions m×n, where each row vector represents an n-dimensional feature vector selected from a functional magnetic resonance imaging (fMRI) whole-brain image dataset; Train These are the corresponding m one-hot true class labels of dimension C, with dimensions m×C, and their row vectors representing the true class labels of the m training vectors. Accordingly, X Test It is a test data matrix with dimensions (Mm)×n; Y Test This corresponds to the real category label, with dimensions (Mm) × C. For example... Figure 3 As shown below the dashed line, the model is tested every 200 training iterations to verify whether it has converged.

[0087] y = X·w + b (1)

[0088] The variational correlation assessment classification model provided in this application is mainly implemented through two parallel fully connected layers to establish the input data X(XTrain or X Test ) and predicted label y(y Train or y Test The linear regression relationship between ); in formula (1), · represents the dot product of two matrices, and w and b are the weight matrix and bias matrix, respectively. Naive Bayes theory assumes the continuity of variables and the independence between variables in the approximate model. In the variational correlation assessment classification model, for m training samples and their corresponding true labels, {(X1,Y1),…,(X m ,Y m There exists a weight vector w between the data matrix and its true labels that accurately represents the linear relationship between them. True And the elements within it are independent of each other, w True The joint posterior distribution expression is shown in Equation (2).

[0089]

[0090]

[0091] Because the true joint posterior distribution function P(w) is accurately obtained True |X Train ,Y Train The specific expression of this distribution is extremely complex. Therefore, this application introduces a variational Bayesian method to approximate the true posterior distribution. The variational Bayesian method assumes a posterior distribution function Q(w|X). Train ,y Train The approximate target posterior distribution is given by formula (3), where y Train This corresponds to the predicted category label of the training data. Formula (4) is the Bayesian expression function of the true posterior distribution, while formula (5) is the central isotropic Gaussian joint prior distribution function of the weight matrix w. This non-information prior distribution is a Gaussian distribution defined by two hyperparameters, namely the prior average. and prior variance (This application simplifies the two hyperparameters to) (i.e., standard Gaussian distribution); for these two hyperparameters, if there is relevant prior information, the convergence of the model can be accelerated by presetting the two hyperparameters.

[0092]

[0093]

[0094]

[0095] To effectively assess the similarity between the true joint posterior distribution and the constructed posterior distribution, the Kullback-Leibler divergence is introduced, as shown in Equation (6). As mentioned earlier, the greater the similarity between the two distributions, the smaller the KL divergence. After replacing the true posterior distribution in Equation (6) with the Bayesian equation in Equation (4), the KL divergence becomes the prototype of the model loss equation, as shown on the right-hand side of Equation (6). It can be seen that the KL divergence mainly depends on the constructed posterior distribution Q(w|X). Train ,y Train The KL divergence between the joint prior distribution P(w) and the joint prior distribution P(w), i.e., D KL [Q(w|X Train ,y Train )||P(w)], which is the second part on the right side of formula (6); and the last part on the right side of formula (6), which is in w~Q(w|X Train ,y Train Under the condition log[P(y)] Train |X Train ,w)] -1 The expected value. When the data and its corresponding true category labels are determined, the first part of the right side of formula (6) is log[P(Y Train |X Train )] is a constant.

[0096] To minimize the KL divergence expressed by Equation (6), an appropriate constructive posterior distribution Q(w|X) must be defined. Train ,y Train As shown in Equation (7), the loss equation, as an approximate expression of the KL divergence in Equation (6), also contains two parts: For the second part on the right side of Equation (6), combined with the independence assumption of all parameters in the variational Bayes method, Equation (7) expresses it as the sum of the KL divergences between the posterior and prior distributions corresponding to a single weight value ω. In model training, this term continuously regularizes the posterior distribution to make it close to the preset prior distribution; The last part on the right side of Equation (6) is expressed in Equation (7) as a certain distance (Euclidean distance in this application). In model training, this term regularizes the predicted label to make it as similar as possible to the true label.

[0097]

[0098]

[0099] During training, in order to find the local minimum point of the loss equation, the two parts on the right side of the equation (7) need to decrease simultaneously, but the directions of change of these two parts are opposite, such as... Figure 4 As shown in (a).

[0100] It should be noted that the variational correlation assessment classification model provided in this application includes an average layer and a variance layer. The average layer and the variance layer are two parallel fully connected layers, used to determine the μ value and σ value for each ω, respectively. 2 The value. Therefore, in formula (8), the final predicted label y... Train The description states that the predicted label is the average predicted label y. μ and variance prediction labels The sum of the square roots of .

[0101] Specifically, the averaging layer evaluates the correlation between weight values ​​and specific experimental conditions, taking into account the interaction effects of multiple variables, while the variance layer evaluates the robustness of the averaging layer after applying noise. Considering the differentiability of two parallel fully connected layers when using stochastic gradient descent (SDG) during model training, a "data resampling" method is introduced: a data resampling layer is designed with y... μ and Gaussian resampling matrices of the same dimension have elements that all follow a standard Gaussian distribution θ ~ N(0,1), where ⊙ in equation (8) represents the Hadamard product. By introducing the resampling matrix, the construction of a Gaussian distribution N(μ,σ) is achieved. 2 Sampling of w in ) allows for the prediction of the label y. Train Sampling is performed to evaluate the predicted label y. Train Compared with the real label Y Train The distance between them is shown in the second part on the right side of the equal sign in formula (7).

[0102] In some embodiments, determining whether the trained variational correlation assessment classification model has converged includes:

[0103] Determine whether the number of training iterations in a single iteration is greater than or equal to a preset number.

[0104] If so, perform a model convergence test. If the classification accuracy of the test set is higher than the preset accuracy, the model is determined to be converged; otherwise, the model is determined not to be converged.

[0105] If the model fails to converge, iterative training continues. If the model still fails to converge after a preset number of training iterations, the current iteration ends, and the number of features required for the next iteration is increased.

[0106] like Figure 3As shown, during model training and testing, independent test data is introduced to evaluate classification accuracy after a certain number of training iterations (one test every 200 training iterations). To ensure model convergence, each model is trained for 1000 iterations by default, followed by a model convergence test (whether the classification accuracy of the test dataset is higher than the random accuracy). If the test determines that the model has not converged, another 1000 training cycles are completed, followed by another test to see if the model has converged. Each feature extraction calculation iteration is trained for a maximum of 3000 iterations. If the model still has not converged after 3000 iterations, the current calculation iteration is interrupted. The system determines that the model may not converge due to insufficient effective information of the current features, and then exits the current calculation iteration, increasing the number of features required for the next calculation iteration. Specifically, in the calculation data selection stage of the next iteration, `max_fea` is modified to increase the number of features calculated simultaneously. That is, in the next calculation iteration, all features from the current iteration are retained, and some unresolved new features are added to participate in the calculation. Once the model is determined to have converged after 1000, 2000, or 3000 training cycles, the next stage of the computational iteration—effective feature extraction—begins.

[0107] In some embodiments, the i-th iteration position sequence is composed of a sub-position sequence ζ used to locate effective feature positions in the local dimension. i Composed of supplementary sequences;

[0108] The supplementary sequence is used to convert the local dimension of the sub-position sequence into the global dimension of the position sequence.

[0109] The weights of the two parallel linear regression layers in the classification model, the weight matrix of the mean layer, and the weight matrix of the variance layer together describe the constructed posterior distribution of each independent weight, i.e., ω[i,j]~N(μ[i,j],σ). 2 [i,j]), where μ[i,j] represents the value of the average layer weight matrix at position [i,j], and the dimension parameter of both weight matrices is [n*C]. For example... Figure 5 As shown, each column of the weight matrix is ​​associated with a specific stimulus condition; based on the goal of extracting the fewest and most effective features to ensure algorithm convergence, the following settings are configured: and This serves as the criterion for effective feature selection. All voxel features need to be evaluated for effectiveness one by one. If, after model convergence, the distribution hyperparameter corresponding to a certain weight ω[i,j] simultaneously satisfies... and Then the corresponding element ρ in the "position selection matrix" i [i,j] equals zero. Then, the "position selection matrix" is averaged column-wise to obtain the "position selection sequence" λ. i When the "position selection sequence" λ iWhen an element at a certain position in the sequence is equal to zero, the sparse position selection sequence γ is... i The corresponding element at the specified position is equal to zero; conversely, when the "position selection sequence" λ i When an element at a certain position is not equal to zero, a sparse position selection sequence γ is formed. i The corresponding element in the sequence is equal to 1. That is, if the constructed posterior distribution of the weights for a feature under different stimulus conditions is concentrated around some meaningless points, then that feature point will no longer participate in subsequent computational iterations. "Sparse position selection sequence" γ i This describes the effectiveness evaluation position of each feature after completing the "i-th model training and testing" process, where all features determined in the "i-th computational data selection" are used; γ i It needs to be converted into a "backfill position sequence" τ of shape [K*1]. i (K = n + g, where g is the number of meaningless feature positions determined by all previous computational iterations). The transformation from the "sparse position selection sequence" to the "backfill position sequence" is mainly achieved based on the "position sequence" used in the "i-th computational data selection" process. The "backfill position sequence" describes the positions of all voxel features participating in the "i-th computational iteration" in all feature dimensions involved in the completed computational iterations. Finally, since some voxel features were judged as meaningless in the completed "i-th computational iteration," in order to ensure the computational efficiency of the next computational iteration (ensuring that the number of features participating in each computational iteration is equal to max_fea), l position sequences that have not participated in previous computational iterations to evaluate new features are constructed to supplement max_fea (since all features supplementing max_fea must participate in the next computational iteration, each element of this position sequence is equal to 1). Then, the "backfill sequence" is concatenated with this position sequence to obtain the new "(i+1)-th sub-position sequence" ζ. i+1 Its dimension is L = n + g + l, where l equals the number of features judged as invalid in the "i-th feature selection" stage. In summary, the vector "(i+1)-th sub-position sequence" ζ i+1 It consists of the following parts: the feature positions judged as valid in the "i-th valid feature extraction" stage, the supplementary element positions used to supplement max_fea, and all feature positions judged as invalid in the "i-th valid feature extraction" stage and the previous (i-1) valid feature extraction stages.

[0110] like Figure 4 As shown in (b), the reason why the effective feature extraction stage can achieve accurate identification of effective features is that the default distribution depicted by the curve is precisely the default posterior distribution Q(w|X) of each weight value w. Train ,y TrainThat is, in each iteration of calculation, the weight values ​​are pre-defined to conform to the same distribution. Figure 4 (b) The prior distribution curve depicts the prior distribution of the weights, i.e. Meanwhile, the prior distribution is also a crucial part of the first adversarial strategy in the preceding "model training and testing" phase; and Figure 4 (b) depicts the feature selection quantification standard curve, which represents the feature evaluation quantification standard defined in the effective feature extraction stage. Regarding the above three distributions, the Gaussian distribution is consistently selected in this application, and the hyperparameters of the three Gaussian distributions must satisfy formula (9):

[0111]

[0112] The variational correlation assessment classifies the classification model, ensuring that the predicted labels are as consistent as possible with the true labels during training and testing; the constructed posterior distribution of ω is also designed to be as consistent as possible with its prior distribution. In fact, after model convergence, the vast majority of the constructed posterior distributions of ω closely approximate the prior distribution (concentrated near zero), with only a small portion of the constructed posterior distributions of ω concentrated in certain important intervals (with sufficiently large central values ​​and sufficiently small variances). This approach utilizes effective feature evaluation criteria and two different regularization directions of the prior distribution to identify and confirm features distributed within important intervals.

[0113] A complete computational iteration includes three stages: computational data selection, model training and testing, and effective feature extraction, such as... Figure 2 As shown in the diagram. The end of one computational iteration confirms the sub-position sequence required for the next computational iteration and also initiates the next computational iteration. The number of iterations required for the entire model to eventually converge depends on the data supplementation efficiency at the beginning of each new computational iteration: when the number of features that have not participated in the computational iteration is sufficient to supplement max_fea, the number of features n participating in one iteration is equal to max_fea; otherwise, n is equal to the number of effective features obtained after the previous iteration.

[0114] It should be noted that the analysis is not finished after all voxel features of the whole brain have undergone one feature calculation iteration; all valid features continue to participate in the calculation iteration until the sparse position selection sequence determined by this calculation iteration is completely consistent with the prepared sub-position sequence for the next iteration (or the position sequence of this iteration is completely identical to the sparse position selection sequence). At this point, the validity analysis of all features of the whole brain is concluded. The goal is to select the fewest, and also the most effective, features from all voxels of the whole brain, while ensuring model convergence.

[0115] S104, determine each effective feature location sequence, convert the effective feature location sequence into a three-dimensional brain structure matrix, and overlay the three-dimensional brain structure matrix onto a preset standard human brain template to identify effective features related to each stimulus condition.

[0116] Specifically, such as Figure 6 As shown, the "location selection matrix" of the last calculation iteration, where each column corresponds to the effective feature location sequence related to a specific experimental condition (column C of the "location selection matrix" represents a total of C stimulus conditions), can be reshaped to the same shape as the input whole-brain data matrix (in this application, the shape size is 79×95×79), i.e., the effective feature location whole-brain matrix. Then, the effective feature location whole-brain matrix under each condition can be overlaid onto the standard template to obtain the activation location topology map for that stimulus condition.

[0117] In some embodiments, it also includes:

[0118] The target category is input into the variational correlation assessment classification model to obtain the prediction vector;

[0119] The sub-correlation coefficient is calculated based on the prediction vector; wherein the sub-correlation coefficient is the percentage of the difference between each voxel and the linear result calculated with the weight values ​​corresponding to different categories in the two parallel regression layers in the difference between the first element and any other element of the prediction vector.

[0120] The cognitive state matrix is ​​obtained based on the sub-correlation coefficients; the cognitive state matrix is ​​used to represent the time-domain changes of whole-brain voxels.

[0121] Specifically, as Figure 7 (a) The right-hand side of the equation illustrates the process of predicting the category label for a specific N-dimensional whole-brain data point belonging to category _1 after model convergence. If the model can accurately predict the stimulus condition label associated with this brain signal, then the predicted signal for the category _1 stimulus-related functional magnetic resonance imaging (fMRI) whole-brain image signal is a C-dimensional vector, and its first element (marked with an asterisk) is also the largest element (one-hot label) among all elements in the vector. For example... Figure 7 As shown in (b), the sub-correlation coefficient is the sum of the prediction vector elements represented by the target category ( Figure 7 The first element of the prediction vector in (a) ) and any other element ( Figure 7 (a) The percentage of the difference between each voxel and the regression result calculated from the weights of different categories in the two parallel regression layers in the difference of other elements in the prediction vector, i.e., formula (10). For feature x iIn formula (10), the two parts on the right-hand side represent the mean layer and the variance layer, respectively. i The classification contribution metric results, where y μ_1 It is the predicted label element corresponding to category _1 calculated by the average layer, b m_1 It is the bias value of the average layer related to category _1. It is the feature x in the average layer i The weight value corresponding to category _1. For a C-class classification model, there are a total of (C-1) sub-correlation coefficients for a single voxel. The average of all sub-correlation coefficients is the correlation coefficient between the voxel and the target stimulus condition, as shown in formula (11). The sum of the correlation coefficients of all effective voxel features in the whole brain is equal to 1; the correlation coefficient of a voxel cluster is the sum of the correlation coefficients of all voxels in the voxel cluster. By arranging the correlation coefficients of a feature with a specific stimulus condition in sequence according to the time-domain sequence of different volume data collection, the cognitive state dynamics of the feature under this stimulus condition can be obtained, such as Figure 7 As shown in (c), the correlation coefficients of all effective voxels constitute the cognitive state matrix at the time point of the volume signal acquisition. Arranging the cognitive state matrices corresponding to multiple volumes acquired during the execution of a specific task in sequence yields the cognitive state dynamics under that stimulus condition.

[0122] By overlaying the cognitive state matrix onto a standard human brain template, the topological distribution of different correlation coefficients can be expanded.

[0123]

[0124]

[0125] The stability and robustness of the model's classification performance are typically achieved through cross-validation. This process can be repeated multiple times to ensure that each sample in the database has an equal chance of being selected as test or training data. For each whole-brain BOLD signal, after each convergence calculation, a cognitive state matrix is ​​generated, consisting of the correlation coefficients of all effective voxels in the whole brain, determined by a variational correlation evaluation algorithm. Considering the potential impact of different training / test dataset partitions on the cognitive state matrix, and the assumption that each functional magnetic resonance imaging signal acquired from the same task stimulus block may represent a unique cognitive state at a specific acquisition time, all training / test dataset partitioning methods were considered in the analysis. For example, if a total of 8 whole-brain signals were acquired in a stimulus block, and a training / test dataset partitioning ratio of 80% / 20% was used, there would be 28 (…). There are 4 different training / test dataset splitting methods, which also represents the number of variational correlation assessment classification model sets that need to be constructed. Finally, the cognitive state matrix corresponding to a certain whole-brain BOLD signal is the average of all cognitive state matrices obtained by the various classification models associated with it after convergence on their respective datasets.

[0126] This application also employs hierarchical clustering for secondary feature extraction. Hierarchical clustering is a type of clustering algorithm that creates a hierarchical nested clustering tree by calculating the similarity between data points of different categories. In the clustering tree, the original data points of different categories form the bottom layer, and the top layer is the root node of a cluster. There are two methods for creating clustering trees: bottom-up merging and top-down splitting. This application involves the merging method.

[0127] Hierarchical clustering merging algorithms calculate the similarity between two classes of data points, combining the two most similar data points among all data points, and iterating this process repeatedly. Simply put, hierarchical clustering merging algorithms determine the similarity between data points in each class and all other data points by calculating the distance between them; the smaller the distance, the higher the similarity. The two closest data points or classes are then combined to generate a cluster tree. There are three methods for calculating the distance between two combined data points: Single Linkage, Complete Linkage, and Average Linkage. Below, we will introduce each of these three methods and their respective advantages and disadvantages.

[0128] The Single Linkage method calculates the distance between two data points in a combination by using the distance between the two closest data points. This method is susceptible to extreme values. Two very different data point sets might be grouped together because two extreme data points are close to each other. Conversely, the Complete Linkage method uses the distance between the two farthest data points in a combination as the distance between the two data point sets. Its disadvantage, also opposite to Single Linkage, is that two similar data point sets might not be grouped together because two extreme data points are far apart. The Average Linkage method calculates the distance between each data point in one combination and all data points in another combination, and uses the average of all distances as the distance between the two data point combinations. This method involves more computation but is more reasonable than the previous two methods, and therefore is the most widely used.

[0129] As the correlation coefficient is explained, the sum of the correlation coefficients of all effective voxel features in the whole brain at a given data acquisition time point equals 1. Based on the assumption that each whole-brain data point may represent a unique cognitive state, the cognitive state dynamics of a specific cognitive task are composed of the cognitive state matrices of each of the multiple consecutive BOLD signal acquisition time points during the task stimulus or execution. It is also possible to study the cognitive state dynamics of a single voxel with different cognitive associations in the time domain. By analyzing the different cognitive state dynamics of voxel features, the cognitive state dynamics of different activated brain regions related to a certain cognitive function can be determined.

[0130] Given the need to generate a cognitive state matrix at each data acquisition time point, the quantitative contribution of each voxel cluster to accurate classification at each data acquisition time point can be obtained by summing the correlation coefficients (RI) of all voxels in each voxel cluster (i.e., the correlation coefficient of each voxel cluster, representing the sum of the classification contributions of all voxels in that voxel cluster at that time point). The classification contributions of each voxel cluster during the cognitive task are arranged sequentially according to the data acquisition time series, and the change process of the quantified results in the time domain is defined as the cognitive state dynamic of that voxel cluster. It is important to emphasize that although the activated features cluster into multiple voxel clusters after effective feature identification in the variational correlation evaluation model, this does not mean that the temporal change dynamics of the correlation coefficients of all voxels within a single voxel cluster follow the same pattern. The next step is to use the cognitive state dynamics of all voxels in a voxel cluster as the analysis object, and apply a two-level hierarchical clustering method to identify the cognitive state dynamics of that voxel cluster that conforms to the temporal change dynamics of the correlation coefficients of most voxels, and define this as the dominant cognitive state dynamic of that voxel cluster. It should be noted that the cognitive state dynamics mentioned in this application specifically refer to the dominant cognitive state dynamics that are compatible with most voxel features.

[0131] The variational correlation assessment method provided in this application quantifies the contribution of individual voxels in the execution of specific cognitive functions. Compared with traditional computational neuroscience models, it eliminates the need for feature pre-selection before model training, allowing the whole brain to be treated as a whole and simultaneously analyzing the features of all brain voxels. The batch feature input algorithm cleverly strikes a balance between ensuring computational efficiency and avoiding overfitting. Two sets of adversarial strategies, implemented by two parallel linear regression layers, ultimately enable the simultaneous execution of feature selection and model training. Each feature selection iteration requires determining the features participating in the current iteration based on the effective feature position sequence obtained from the previous iteration. Finally, the model converges by identifying and extracting the feature with the smallest volume that best represents the target stimulus condition.

[0132] As a specific implementation method, this application verifies the effectiveness of the variational correlation assessment classification model in feature extraction applied to functional magnetic resonance imaging data analysis. Based on the visual stimulus content, a six-classification model under the variational correlation assessment algorithm architecture is established. The distribution location of various stimulus activation voxels extracted by the variational correlation assessment classification model is verified to be consistent with the anatomical location of brain regions activated by higher visual stimuli as previously reported. The "operational" dynamics and "representational" locations of the brain during the recognition and cognition process of different categories of higher visual stimuli are analyzed.

[0133] The experiment recruited 15 visually normal participants (including 6 women, right-handed), with a mean age of 29.4 years (participant age range: 21-29 years). Data acquisition was performed using a 3Tesla (T2*-weighted echo planner imaging, EPI, TR / TE = 2000 / 30ms, FA = 90°, 1943Hz / px bandwidth, 80×80 matrix, in-plane dimension = 3×3mm, field-of-view (FoV) = 240mm) Philips Achieva dStream MRI scanner equipped with a 32-channel head coil. The acquisition range covered the entire brain.

[0134] Visual stimuli were projected onto a screen located in the MRI scanning chamber via an LCD projector (JVC DLA RS66E, JVC Ltd.). The projector was connected via a DVI extension system to a computer in the control room used to control the execution of the stimulus paradigm. The screen in the scanning chamber used to display the images measured 26.5cm × 21.2cm, with a resolution of 1280 × 1024px and a refresh rate of 60Hz. The subject lay supine in the scanning chamber, and the visual stimuli were projected onto both eyes via a reflective mirror mounted above the head coil, with the viewing distance adjusted to 63cm. The stimulus paradigm was executed using the Psychopy V1.79 application installed on the (Neuro)Debian operating system. The subject's response to the stimulus signal was fed back via a two-button response system and recorded in the work log of the computer in the control room.

[0135] like Figure 8As shown in (a), each participant needed to complete 4 runs; each run contained 12 stimulus blocks, including two blocks of each stimulus lasting 16 seconds (each block played 16 different pictures of the same category), each picture lasting 900ms, with a 100ms interval between pictures. Two blocks of each stimulus category appeared (the two blocks did not appear consecutively), meaning one block of each stimulus category appeared in the first half of the run, and one block of each stimulus category appeared in the second half of the run. During the picture-viewing experiment, participants also needed to pay attention to a randomly occurring one-back task to ensure sustained attention: each block might present two consecutive mirror images (it could appear zero times, once, or twice), and participants were required to press a button to identify the one-back task. A crosshair was placed in the center of the screen to ensure the participant's visual field was centered on the crosshair; this crosshair was green during the stimulus block presentation; there was an 8-second rest period between different picture blocks, during which the crosshair in the visual field turned white, and 1.5 seconds before the next block began, the crosshair in the visual field changed from white to green.

[0136] It should be emphasized that although this experiment involves one-back working memory, the temporal resolution of the data is insufficient to analyze the neural response mechanisms related to working memory because each image appears for only 900ms while the TR time of functional magnetic resonance imaging (fMRI) data acquisition is 2000ms. Therefore, the cognitive process research in this chapter focuses on the high-level visual brain region localization task with different types of visual stimulus blocks as the subjects.

[0137] This experiment used advanced visual stimuli from six different stimulus categories, with 24 grayscale images for each category. The six categories included: face stimuli, human body stimuli with faces removed, artificial tool stimuli, house stimuli, outdoor scenery stimuli (including natural or street scenes), and phase-disrupted image stimuli. Figure 8 As shown in (b). After the original image is converted to grayscale, its resolution needs to be adjusted to 400×400px. Then, the lumMatch application in the SHINE toolbox is used to adjust the brightness of all images to conform to a normal distribution range with a mean of 128 and a standard deviation of 70. The images are presented at an angle of approximately 10°×10°.

[0138] As a specific implementation method, each subject used 384 functional magnetic resonance imaging (fMRI) data points across 48 blocks. 80% of these data were used as the model training dataset, and the remaining 20% ​​as the model testing dataset. Assuming model convergence, the classification accuracy directly reflects the amount of useful information extracted from the BOLD voxel features. Variational correlation assessment analysis, constructed using fMRI data for advanced visual stimulation, showed that all classification models achieved a relatively ideal accuracy (98.6% ± 1.1%, the mean and standard deviation of accuracy for all models) after the effective feature extraction phase, ensuring the effectiveness of the extracted features.

[0139] It should be noted that, for each of the six different stimulus conditions, the classification model needs to construct six different linear regression functions for each calculation iteration; after the effective feature extraction stage is completed, the activation topology map corresponding to each advanced visual stimulus condition can be obtained.

[0140] In some embodiments, such as Figure 9 As shown, this application provides a variational correlation assessment apparatus, including:

[0141] The acquisition module 201 is used to acquire magnetic resonance imaging data and perform preprocessing to obtain data to be processed; wherein, the magnetic resonance imaging data includes brain voxel features;

[0142] Extraction module 202 is used to determine the iteration position sequence of the data to be processed, extract the iteration features of the data to be processed, divide the iteration features by a preset ratio, and obtain a training set and a test set;

[0143] The training module 203 is used to input the training set into the variational correlation assessment classification model for iterative training, and to test the variational correlation assessment classification model after reaching the training number, to determine whether the trained variational correlation assessment classification model has converged, and to extract effective features when it converges.

[0144] The identification module 204 is used to determine the effective feature location sequence related to each stimulus, convert the effective feature location sequence into a three-dimensional brain structure matrix, and overlay the three-dimensional brain structure matrix onto a preset standard human brain template to identify the effective features related to each stimulus condition.

[0145] The working principle of the variational correlation assessment device provided in this application is as follows: the acquisition module 201 acquires magnetic resonance imaging data and performs preprocessing to obtain data to be processed; wherein, the magnetic resonance imaging data includes brain voxel features; the extraction module 202 determines the iterative position sequence of the data to be processed and extracts the iterative features of the data to be processed, and divides the iterative features according to a preset ratio to obtain a training set and a test set; the training module 203 inputs the training set into the variational correlation assessment classification model for iterative training, and tests the variational correlation assessment classification model after reaching the training number, determines whether the trained variational correlation assessment classification model has converged, and extracts effective features when it converges; the identification module 204 determines the effective feature position sequence related to each stimulus, converts the effective feature position sequence into a three-dimensional brain structure matrix, and overlays the three-dimensional brain structure matrix onto a preset standard human brain template to identify the effective features related to each stimulus condition.

[0146] It is understood that the method embodiments provided above correspond to the device embodiments described above, and the specific details can be referred to each other, which will not be repeated here.

[0147] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0148] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0149] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction methods implemented in a process. Figure 1One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0150] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0151] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method of evaluating a variational dependence, characterized by, The method comprises the following steps: acquiring magnetic resonance image data and pre-processing to obtain to-be-processed data; wherein the magnetic resonance image data comprises brain voxel features; determining an iteration position sequence of the to-be-processed data, and extracting iteration features of the to-be-processed data, and dividing the iteration features by a preset ratio to obtain a training set and a test set; inputting the training set into a variational correlation evaluation classification model for iterative training, and testing the variational correlation evaluation classification model after a preset number of training times, judging whether the trained variational correlation evaluation classification model converges, and when it converges, extracting effective features; determining an effective feature position sequence of each stimulus-related effective feature, converting the effective feature position sequence into a three-dimensional brain structure matrix, and covering the three-dimensional brain structure matrix to a preset standard human brain template to identify effective features related to each stimulus condition.

2. The method of claim 1, wherein, The effective feature extraction based on the variational correlation evaluation classification model comprises: determining an effective feature sparse position sequence of the current iteration, and determining a sub-position sequence for calculating iteration features next time; judging whether the effective feature sparse position sequence of the current iteration is the same as the sub-position sequence for calculating iteration features next time; if not, entering the next calculation iteration, otherwise, abandoning the current iteration, increasing the number of feature data participating in the calculation and starting the calculation iteration again.

3. The method of claim 1 or 2, wherein: The i-th sequence of effective feature sparse positions consists of a sub-sequence of positions ζ i and a complementary sequence; the supplementary sequence is used to convert the local dimension of the sub-position sequence to the global dimension of the position sequence.

4. The method of claim 3, wherein, The variational correlation evaluation classification model comprises: an average layer and a variance layer arranged in parallel; the average layer is used to evaluate the correlation between the weight value and the specific experimental condition under the consideration of the interaction effect of multiple variables; the variance layer is used to evaluate the robustness of the average layer.

5. The method of claim 4, wherein, The variational correlation evaluation classification model comprises: where X Train is the training input data, y Train is the predicted class label corresponding to the training data, w and b are the weight matrix and bias value matrix respectively, Q(w|X Train ,y Train ) is the posterior distribution function, P(w) is the joint prior distribution function, is the prior mean parameter, is the prior variance parameter, is the Hadamard product, D KL [Q(w|X Train ,y Train )||P(w)] is the KL divergence.

6. The method of claim 1, wherein, The judgment of whether the trained variational correlation evaluation classification model converges comprises: judging whether the number of training times in one iteration is greater than or equal to a preset number; if yes, performing model convergence testing, if the classification accuracy of the test set is higher than a preset accuracy, determining that the model converges, otherwise, determining that the model does not converge; when it is determined that the model does not converge, the current model is continued to be iteratively trained, if the model still does not converge after a preset number of training, the current iteration is ended, and the number of features required for the next calculation iteration is increased.

7. The method of claim 5, wherein, Further comprising: inputting the target category into the variational correlation evaluation classification model to obtain a prediction vector; calculating a sub-correlation coefficient according to the prediction vector; wherein the sub-correlation coefficient is the percentage of the difference between each voxel and the linear result calculated from the weight values corresponding to two parallel regression layers of different categories in the difference between the first element and any other element of the prediction vector; obtaining a cognitive state matrix according to the sub-correlation coefficient; the cognitive state matrix is used to represent the time domain changes of the whole brain voxel specificity.

8. The method of claim 1, wherein, The pre-processing comprises pre-pre-processing and post-pre-processing; the pre-pre-processing comprises image inter-layer time correction, head motion correction, image spatial registration and spatial standardization; the post-pre-processing comprises noise reduction and normalization.

9. The method of claim 1, wherein, a batch feature input method is used to input whole brain voxel features participating in the calculation iteration in batches and sequentially participate in the feature extraction calculation iteration; the iteration position sequence is a sparse sequence.

10. A variational dependence evaluation apparatus characterized by comprising: comprises: an acquisition module configured to acquire magnetic resonance image data, perform preprocessing, and obtain to-be-processed data; wherein the magnetic resonance image data comprises brain voxel features; an extraction module configured to determine an iteration position sequence of the to-be-processed data, extract iteration features of the to-be-processed data, divide the iteration features at a preset ratio, and obtain a training set and a test set; a training module configured to input the training set to a variational correlation evaluation classification model for iterative training, test the variational correlation evaluation classification model after a training number is reached, determine whether the trained variational correlation evaluation classification model converges, and perform effective feature extraction when the trained variational correlation evaluation classification model converges; an identification module configured to determine each extracted effective feature position sequence related to stimulation, convert the effective feature position sequence into a three-dimensional brain structure matrix, overlay the three-dimensional brain structure matrix to a preset standard human brain template, and identify effective features related to each stimulation condition.

Images