A task-state fMRI encoding method, system, device, medium and product under natural stimulation conditions

Abstract

This application discloses a task-oriented fMRI encoding method, system, device, medium, and product under natural stimulus conditions, relating to the field of fMRI encoding. The method includes: acquiring task-oriented fMRI data collected under natural stimulus conditions; constructing an fMRI data matrix; constructing a feature subspace; and constructing a banded ridge regression encoding model. Based on the overall regularization parameter candidate set, the optimal hyperparameters of the banded ridge regression encoding model are determined. A priori voxel spatiotemporal homogeneity under natural stimulus conditions is introduced, and the temporal series correlation coefficient between each voxel and its neighboring voxels is calculated to determine the global regularization parameter after spatiotemporal constraints, and the subspace banded ridge regularization parameter is determined. The subspace banded ridge regularization parameter is substituted into and the banded ridge regression encoding model is solved to determine the regression weights of each feature subspace, thereby realizing task-oriented fMRI encoding for brain region activity analysis. This application alleviates the hyperparameter fluctuation problem in the natural stimulus encoding process.

Description

Technical Field

[0001] This application relates to the field of functional magnetic resonance imaging (fMRI) coding, and in particular to a task-oriented fMRI coding method, system, device, medium, and product under natural stimulus conditions. Background Technology

[0002] In recent years, numerous studies under the natural stimulus paradigm have revealed the brain's coding mechanisms to some extent. However, with the increase in the dimensionality and number of encoded features, strong correlation coefficients are generally found between different features. Traditional ridge regression, based on the assumption of spherical symmetry prior, is no longer able to effectively distinguish the contributions of each subspace. Tour et al. proposed the banded ridge regression method, which introduces an independent regularization intensity for each feature subspace, equivalent to imposing an "anisotropic" banded prior on the feature space. This method can differentiate constraints on the scale and redundancy of different feature subspaces, making it more suitable for handling high-dimensional, multi-subspace, and collinear natural stimulus coding problems.

[0003] However, in practical fMRI modeling, especially when the time length of a single voxel is limited and the signal-to-noise ratio is low, the optimal constraint hyperparameters given by cross-validation are... It is often unstable and easily fluctuates greatly due to accidental noise or differences in partitioning, resulting in a "disorderly" distribution of hyperparameters within the same functional area, which in turn affects the model's prediction accuracy and spatial interpretability. Summary of the Invention

[0004] The purpose of this application is to provide a task-oriented fMRI coding method, system, device, medium, and product under natural stimulus conditions to solve the problem of large hyperparameter fluctuations in the natural stimulus coding process.

[0005] To achieve the above objectives, this application provides the following solution.

[0006] In a first aspect, this application provides a task-state fMRI coding method under natural stimulus conditions, comprising the following steps.

[0007] Acquire task-state fMRI data collected under natural stimulus conditions and construct an fMRI data matrix.

[0008] A feature subspace is constructed based on the fMRI data matrix; each feature subspace corresponds to a type of feature; the features include phoneme vector features, phoneme counting features, word counting features, and contextual semantic features of the large language model.

[0009] A ridge regression coding model is constructed based on the feature subspace and the fMRI data matrix.

[0010] The optimal hyperparameters of the banded ridge regression coding model are determined based on the overall candidate set of regularization parameters; the optimal hyperparameters include the optimal subspace kernel weights and the corresponding initial global regularization parameters.

[0011] Introducing a priori information on the spatiotemporal homogeneity of voxels under natural stimulus conditions, the time series correlation coefficient between each voxel and its neighboring voxels is calculated, and the global regularization parameter after spatiotemporal constraints is determined based on the initial global regularization parameter.

[0012] The subspace banded ridge regularization parameter is determined based on the optimal subspace kernel weight and the global regularization parameter after the spatiotemporal constraints.

[0013] The regularization parameters of the subspace band ridge are substituted into and the band ridge regression coding model is solved to determine the regression weights of each feature subspace, thereby realizing task-state fMRI coding for brain region activity analysis.

[0014] Secondly, this application provides a task-oriented fMRI coding system under natural stimulus conditions, including the following modules.

[0015] The fMRI data matrix construction module is used to acquire task-state fMRI data collected under natural stimulus conditions and construct an fMRI data matrix.

[0016] The feature subspace construction module is used to construct feature subspaces based on the fMRI data matrix; each feature subspace corresponds to a type of feature; the features include phoneme vector features, phoneme counting features, word counting features, and contextual semantic features of the large language model.

[0017] The ridge regression coding model construction module is used to construct a ridge regression coding model based on the feature subspace and the fMRI data matrix.

[0018] The optimal hyperparameter determination module is used to determine the optimal hyperparameters of the banded ridge regression coding model based on the overall regularization parameter candidate set; the optimal hyperparameters include the optimal subspace kernel weights and the corresponding initial global regularization parameters.

[0019] The spatiotemporal homogeneity constraint module is used to introduce the spatiotemporal homogeneity prior of voxels under natural stimulus conditions, calculate the time series correlation coefficient between each voxel and its neighboring voxels, and determine the global regularization parameter after spatiotemporal constraint based on the initial global regularization parameter.

[0020] The subspace banded ridge regularization parameter determination module is used to determine the subspace banded ridge regularization parameter based on the optimal subspace kernel weight and the global regularization parameter after the spatiotemporal constraints.

[0021] The task-oriented fMRI encoding module is used to substitute the regularization parameters of the subspace band ridge into and solve the band ridge regression encoding model, determine the regression weights of each feature subspace, and realize task-oriented fMRI encoding for brain region activity analysis.

[0022] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described task-state fMRI coding method under natural stimulus conditions.

[0023] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described task-state fMRI coding method under natural stimulus conditions.

[0024] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the above-described task-state fMRI coding method under natural stimulus conditions.

[0025] According to the specific embodiments provided in this application, this application has the following technical effects.

[0026] This application constructs an fMRI data matrix and feature subspaces to build a band ridge regression coding model. It introduces a priori voxel spatiotemporal homogeneity under natural stimulus conditions to constrain the overall regularization parameters. The time-series correlation coefficients between each voxel and its neighboring voxels are calculated. Based on the initial global regularization parameters, the spatiotemporally constrained global regularization parameters are determined. This application utilizes the similarity coefficients of neighboring voxels on a three-dimensional spatial voxel network and their time series under natural stimuli to spatiotemporally smooth and correct the initial global regularization parameters, making hyperparameter estimation more robust. This alleviates hyperparameter fluctuations during natural stimulus coding and avoids "chaotic" hyperparameter distribution within the same functional region. Finally, the subspace band ridge regularization parameters are determined based on the optimal subspace kernel weights and the spatiotemporally constrained global regularization parameters. These parameters are then substituted into and solved in the band ridge regression coding model to determine the regression weights of each feature subspace, achieving task-state fMRI coding for brain region activity analysis. This improves the prediction accuracy and spatial interpretability of the band ridge regression coding model. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a schematic flowchart of a task-oriented fMRI coding method under natural stimulus conditions, provided as an embodiment of this application.

[0029] Figure 2 This is an architectural diagram of a task-state fMRI coding method under natural stimulus conditions, provided as an embodiment of this application.

[0030] Figure 3 This is a schematic diagram of the internal calculation process of the spatiotemporal homogeneity constraint module provided in an embodiment of this application. Detailed Implementation

[0031] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0032] To make the objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0033] like Figures 1-2 As shown, this application provides a task-oriented fMRI coding method under natural stimulus conditions, including:

[0034] S1: Acquire task-state fMRI data collected under natural stimulus conditions and construct an fMRI data matrix.

[0035] S2: Construct a feature subspace based on the fMRI data matrix; each feature subspace corresponds to a type of feature; the features include phoneme vector features, phoneme counting features, word counting features, and contextual semantic features of the large language model.

[0036] S3: Construct a banded ridge regression coding model based on the feature subspace and the fMRI data matrix.

[0037] S4: Determine the optimal hyperparameters of the banded ridge regression coding model based on the overall regularization parameter candidate set; the optimal hyperparameters include the optimal subspace kernel weights and the corresponding initial global regularization parameters.

[0038] S5: Introduce the prior of spatiotemporal homogeneity of voxels under natural stimulus conditions, calculate the time series correlation coefficient between each voxel and its neighboring voxels, and determine the global regularization parameter after spatiotemporal constraints based on the initial global regularization parameter.

[0039] S6: Determine the subspace banded ridge regularization parameter based on the optimal subspace kernel weight and the global regularization parameter after the spatiotemporal constraints.

[0040] S7: Substitute the regularization parameters of the subspace band ridge into and solve the band ridge regression coding model to determine the regression weights of each feature subspace, realize task-state fMRI coding, and perform brain region activity analysis.

[0041] In an exemplary embodiment, S1 specifically includes:

[0042] S11: Based on natural stimulus conditions, the BOLD signals of all voxels at all time points are used as task-state fMRI data.

[0043] S12: Construct an fMRI data matrix based on the task-state fMRI data.

[0044] In practical applications, time points are defined as t=1…T, and voxels as v=1…V. The blood-oxygen-level-dependent (BOLD) signals of all voxels at all time points are organized into an fMRI data matrix. Among them, the first fMRI data matrix v Listed as voxels v Time series, , Let T be a T×V dimensional matrix space consisting of real numbers, where T is the number of time points in fMRI and V is the number of voxels. The T-dimensional vector space composed of real numbers is actually the time signal of fMRI.

[0045] Figure 2 middle, X i Let i be the i-th feature subspace, i.e., the stimulus features. Y~Banded Ridge ({ X i}) to be X i and Y Perform zonal ridge regression.

[0046] Based on natural stimuli (such as movies or speech), multiple feature subspaces are constructed on a computer: for example, phoneme vector features, phoneme count features, word count features, and contextual semantic features based on large language models such as GPT-2 or BERT.

[0047] In practical applications, the fMRI data matrix Y is the fMRI signal. Taking the auditory task as an example, if the subject listens to 10 minutes of audio, the collected fMRI signal is the length of 10 minutes of fMRI. The feature subspace is the feature extraction of 10 minutes of audio, such as fundamental frequency, loudness, etc. The length of the extracted features is also 10 minutes. The feature dimensions are different depending on the characteristics of each feature.

[0048] Organize each type of feature into a feature subspace matrix. Where i=1…m represents the i-th feature subspace, and m is the total number of feature subspaces. P i Let be the feature dimension of this feature subspace. Let Y be a T×Pi matrix composed of real numbers. The known quantities include Y and the eigenspace matrix. , The format is used to distinguish between traditional ridge regression and banded ridge regression. This is because traditional ridge regression puts all features together for regularization constraints, while banded ridge regression treats different features as multiple spaces, with each space having independent regularization constraints. The unknowns are the regression weights and regularization parameters that need to be solved later.

[0049] In practical applications, for each voxel v, it is assumed that its time series is generated by a linear combination of its subspaces: ,in, Let be the regression weight of voxel v in the i-th feature subspace. This is the noise term.

[0050] The objective function of the ridge regression coding model is: ,in, >0 is the banded ridge regularization parameter of voxel v in the i-th feature subspace, and || ||2 is the L2 norm.

[0051] Subsequently, a kernel matrix is ​​constructed for each feature subspace. : Introducing regularization parameters >0 and subspace kernel weights ≥0, satisfying: ,in, Let be a T×T dimensional matrix space consisting of real numbers.

[0052] The combined nucleus of construct voxel v: In the combination of voxel v nuclei Define the objective function for multi-kernel ridge regression: The difference between this and the objective function of the aforementioned banded ridge regression coding model lies in the fact that: "banded" refers to multiple spatially independent regularization constraints, while "multi-kernel" here is actually an equivalent form for high-dimensional solutions. Similar to the banded model, the features are first transformed by a kernel function before undergoing multiple spatially independent regularization constraints in the banded ridge regression. These are the dual coefficients. By transforming the relation, we can... and , Establish corresponding relationships: .

[0053] In one exemplary embodiment, S4 specifically includes:

[0054] S41: Based on the overall regularization parameter candidate set, set multiple sets of candidate kernel weight vectors, and construct a combined kernel for each candidate kernel weight vector.

[0055] S42: Based on the combined kernel, perform N-fold cross-validation on each voxel and search for the optimal global regularization parameter in the candidate set of global regularization parameters.

[0056] S43: Based on the optimal global regularization parameter corresponding to all candidate kernel weight vectors, compare the overall performance of all candidate kernel weight vectors, and determine the optimal subspace kernel weight and the corresponding initial global regularization parameter.

[0057] In practical applications, first define the candidate set of overall regularization parameters: , For the first A number of candidate global regularization parameters ∈[1,L], where L is the total number of candidate global regularization parameters.

[0058] Set multiple sets of candidate kernel weight vectors ,in: , Let be the m-th candidate kernel weight vector; K is the total number of candidate kernel weight vectors. For each candidate kernel weight vector... Constructing a combined core : , Let be the candidate kernel weight vector for the i-th feature subspace.

[0059] In a given Under the given conditions, N-fold cross-validation is used for each voxel v, and the optimal global regularization parameter is searched in M, which is the average value of voxel v at each fold. for: ;in, For voxels vIn the k The candidate kernel weight vector, the first The candidate global regularization parameter and the first r The performance evaluation values ​​obtained under cross-validation. r For the first r Cross-validation k For the first k A candidate kernel weight vector.

[0060] For each voxel, the optimal value is obtained on M, and then all candidate kernel weight vectors are compared. Based on the overall performance, select the optimal subspace kernel weight with the best performance. and the corresponding initial global regularization parameters .

[0061] In an exemplary embodiment, S5 specifically includes:

[0062] S51: Introduce the prior of spatiotemporal homogeneity of voxels under natural stimulus conditions, and calculate the time series correlation coefficient between each voxel and its neighboring voxels on a three-dimensional spatial voxel network.

[0063] S52: Define spatial weights based on spatial distance, and determine comprehensive weights by combining the spatial weights with the time series correlation coefficients.

[0064] S53: For each voxel, determine the global regularization parameters after spatiotemporal constraints based on the comprehensive weights and the initial global regularization parameters.

[0065] In practical applications, assume the three-dimensional coordinates of voxel v are... Then the 3×3×3 spatial neighborhood of voxel v Recorded as: ,in, For all satisfied The set of voxel indices u, where u is a neighborhood voxel of voxel v. Neighboring voxels u The three-dimensional coordinates.

[0066] For any Calculate the correlation coefficient of time series : .

[0067] in, and The time series mean of voxels v and u; The time series vector of voxel u; The observed value of the BOLD signal for voxel u at time point t; The value of the BOLD signal observed for voxel v at time point t is given.

[0068] Gaussian spatial weights are defined based on spatial distance. : .

[0069] in, >0 represents the preset spatial scale parameter.

[0070] Temporal correlation and spatial distance are combined into a comprehensive weight. : .

[0071] For each voxel v, based on the initial global regularization parameters of the voxels in the neighborhood. We perform a weighted average to obtain the global regularization parameters after spatiotemporal constraints. : ,in, >0 is a very small constant to prevent the denominator from being zero.

[0072] In obtaining Then, the final subspace banded ridge regularization parameters are calculated based on the equivalence relation: , Let be the candidate kernel weight vector for the i-th feature subspace of voxel v; then, for each voxel v, solve the banded ridge regression problem in the original feature space: And obtain the regression weights of each feature subspace. An analytical solution can also be used: Where X is the feature subspace. , This is a block-diagonal regularization matrix. , The ridge regularization intensity (penalty coefficient) applied to the i-th feature subspace of voxel v, i.e., the subspace banded ridge regularization parameter. Let i be the p×p×p identity matrix, i∈[1,m].

[0073] In practical applications, the predicted fMRI signal under a given feature is obtained by multiplying the regression weights of each feature subspace by the new feature. Then, Pearson correlation is performed between the predicted fMRI signal and the actual fMRI signal to determine the fMRI brain region corresponding to the feature.

[0074] If the feature is known, such as loudness, the brain regions that respond to sound can be determined. Higher-level features have a greater impact, especially abstract features, such as social attributes. This application studies brain activation from this perspective.

[0075] Because fMRI has high noise, the results of solving each voxel fluctuate greatly (the feature prediction results of each voxel are sometimes good and sometimes bad), and the regions are inconsistent. However, the brain has functional homogeneity. It is based on this characteristic that the task-state fMRI coding method under natural stimulus conditions provided in this application is similar to spatial smoothing, which can make the solved hyperparameters more similar in regions, and the region activation is more consistent. This makes it more reliable in analyzing the relationship between features and brain function, and the activation regions are more consistent across groups.

[0076] As a preferred embodiment, this application uses the Pearson correlation coefficient of voxel time series as a measure of temporal similarity, and constructs a Gaussian spatial kernel using three-dimensional Euclidean distance as the weight of the spatial scale, i.e., Gaussian spatial weight.

[0077] As another preferred implementation, the metric used to measure the similarity of voxel time series can also be other measures that can reflect the similarity or predictability of time series, such as: Euclidean distance based on time series differences and converted into similarity weights through monotonic transformation, or R-squared based on the prediction performance of the encoding model. 2 Similarities include cosine similarity and Spearman ranking.

[0078] Similarly, in spatial scale modeling, the Gaussian form of spatial weights is merely a specific weighting function form. Using other spatial weighting functions or neighborhood definitions to embody the idea of ​​"greater weight for closer objects in space" is also within the scope of this application. For example, uniform weights within a fixed window (boxcar kernel), exponentially decaying kernels, Laplacian kernels, and weighting functions based on geodesic distance from the cortical surface rather than Euclidean distance from the voxel grid can be used. As long as a larger smoothing weight is provided for nearby voxels and a smaller weight for distant voxels in the spatial dimension, and the overall regularization parameter is adjusted in conjunction with temporal similarity... Perform weighted smoothing.

[0079] This application provides a task-oriented fMRI coding system under natural stimulus conditions, including:

[0080] The fMRI data matrix construction module is used to acquire task-state fMRI data collected under natural stimulus conditions and construct an fMRI data matrix.

[0081] The feature subspace construction module is used to construct feature subspaces based on the fMRI data matrix; each feature subspace corresponds to a type of feature; the features include phoneme vector features, phoneme counting features, word counting features, and contextual semantic features of the large language model.

[0082] The ridge regression coding model construction module is used to construct a ridge regression coding model based on the feature subspace and the fMRI data matrix.

[0083] The optimal hyperparameter determination module is used to determine the optimal hyperparameters of the banded ridge regression coding model based on the overall regularization parameter candidate set; the optimal hyperparameters include the optimal subspace kernel weights and the corresponding initial global regularization parameters.

[0084] The spatiotemporal homogeneity constraint module is used to introduce the spatiotemporal homogeneity prior of voxels under natural stimulus conditions, calculate the time series correlation coefficient between each voxel and its neighboring voxels, and determine the global regularization parameter after spatiotemporal constraint based on the initial global regularization parameter.

[0085] The subspace banded ridge regularization parameter determination module is used to determine the subspace banded ridge regularization parameter based on the optimal subspace kernel weight and the global regularization parameter after the spatiotemporal constraints.

[0086] The task-oriented fMRI encoding module is used to substitute the regularization parameters of the subspace band ridge into and solve the band ridge regression encoding model, determine the regression weights of each feature subspace, and realize task-oriented fMRI encoding for brain region activity analysis.

[0087] In practical applications, this application, based on existing ridge regression, embeds a global regularization parameter. The structure with spatiotemporal constraints is used, while the remaining parts, such as data reading, feature extraction, kernel construction, and band ridge solving, can all be implemented using existing mature solutions. These are only used as the basic units of the system in this application.

[0088] Specifically, this application includes at least: a ridge regression coding model construction module and a spatiotemporal homogeneity constraint module connected thereto. The output terminal is connected to the input terminal of the spatiotemporal homogeneity constraint module of this application to obtain... It is connected to the regression solution unit in the banded ridge regression coding model construction module to realize the insertion and rewriting of the original hyperparameter flow.

[0089] like Figure 3 As shown, the spatiotemporal homogeneity constraint module is the core structure of the entire device. Internally, it can be subdivided into substructures such as a spatial neighborhood determination unit, a temporal similarity calculation unit, a spatial weight generation unit, and a weighted smoothing unit, which are connected sequentially according to a fixed data flow order. The spatial neighborhood determination unit determines a fixed-size (e.g., 3×3×3) local spatial neighborhood for each voxel based on its coordinate information in the 3D mesh, and passes the neighborhood index to subsequent units. The temporal similarity calculation unit simultaneously receives the fMRI response data from the training set and the neighborhood index, and for each voxel… and its neighboring bodies Calculate time series and The Pearson correlation coefficient, i.e., temporal correlation, is calculated and used as a measure of functional homogeneity. The spatial weight generation unit further calculates Gaussian spatial weights based on the spatial distance between voxels, and multiplies the spatial weights by the temporal similarity point by point to obtain the result for each voxel. Its neighboring voxels The overall weight between The weighted smoothing unit simultaneously receives the overall weight. And the output of the ridge regression coding model construction module We then perform a weighted average to obtain the final smoothing parameter, which is the global regularization parameter after spatiotemporal constraints. and will The output is sent to the ridge regression solution unit in the ridge regression coding model construction module. Through the above structural division and connection relationships, this application adds a new line for the overall regularization parameter without changing the main body of the existing ridge regression modeling process. The spatiotemporal homogeneity constraint path enables smooth estimation of parameters in both spatial and temporal dimensions.

[0090] 1. Wide applicability: This application only applies to the overall regularization parameter. The application adds a spatiotemporal homogeneity constraint module at the layer level. The remaining parts, such as data reading, feature subspace construction, and multi-kernel / band ridge solving, can all use the existing implementation. Therefore, as long as there is a time-aligned task-state fMRI and corresponding feature subspace (including natural film, natural speech and its acoustic, linguistic, visual and other multimodal features), this application can be directly embedded into the original coding model process.

[0091] 2. Spatiotemporal Joint Optimization: Due to the inherent functional homogeneity of fMRI signals, which is even more pronounced in natural stimuli, current methods primarily utilize spatial information. Incorporating temporal similarity into natural stimuli makes the constraint weights more reasonable and efficient.

[0092] 3. Improve the spatial consistency of the model: unstable This can lead to some voxels within the same functional region being weakly regularized or even overfitted, while other voxels are over-regularized. This application enables... Simultaneously relying on multi-voxel information that is "spatially adjacent and has similar temporal responses," rather than solely depending on the limited cross-validation results of a single voxel, thus reducing the impact of noise on the overall regularization parameter. It exhibits a smooth and continuous spatial distribution within functionally homogeneous regions. Techniques directly derived from the spatial neighborhood division within the spatiotemporal homogeneity constraint module, Pearson correlation metric, spatial Gaussian weighting, and neighborhood weighted smoothing enhance the spatial continuity and interpretability of the weight graph.

[0093] 4. Low computational cost: Directly applying spatial regularization to the regression weights requires redesigning a large-scale optimization problem, resulting in high computational costs. This application only addresses... Add a local weighted smoothing operation without adding a new cross-validation dimension or changing the process. The search space is small, and the additional overhead is only the local computations of neighborhood correlation and weighted averaging, which is easy to implement in parallel.

[0094] In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments. The computer device can be a server or a terminal. The computer device includes a processor, a memory, an input / output interface (I / O), and a communication interface. The processor, memory, and I / O interface are connected via a system bus, and the communication interface is connected to the system bus via the I / O interface. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device stores data to be processed. The I / O interface of the computer device is used for exchanging information between the processor and external devices. The communication interface of the computer device is used for communicating with an external terminal via a network connection. When the computer program is executed by the processor, it implements the above-described methods.

[0095] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0096] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0097] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0098] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by hardware related to computer program instructions. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0099] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0100] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0101] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A task-state fMRI coding method under natural stimulus conditions, characterized in that, include: Acquire task-state fMRI data collected under natural stimulus conditions and construct an fMRI data matrix; Construct a feature subspace based on the fMRI data matrix; Each feature subspace corresponds to a type of feature; the features include phoneme vector features, phoneme counting features, word counting features, and contextual semantic features of the large language model; A ridge regression coding model is constructed based on the feature subspace and the fMRI data matrix; The optimal hyperparameters of the banded ridge regression coding model are determined based on the overall candidate set of regularization parameters; the optimal hyperparameters include the optimal subspace kernel weights and the corresponding initial global regularization parameters. Introducing a priori information on the spatiotemporal homogeneity of voxels under natural stimulus conditions, the time-series correlation coefficient between each voxel and its neighboring voxels is calculated. Based on the initial global regularization parameter, the spatiotemporally constrained global regularization parameter is determined, specifically including: Introducing the prior of spatiotemporal homogeneity of voxels under natural stimulus conditions, the temporal correlation coefficient between each voxel and its neighboring voxels is calculated on a three-dimensional spatial voxel network. Spatial weights are defined based on spatial distance, and a comprehensive weight is determined by combining the spatial weights with the time series correlation coefficient. For each voxel, the global regularization parameters after spatiotemporal constraints are determined based on the comprehensive weights and the initial global regularization parameters. The subspace banded ridge regularization parameter is determined based on the optimal subspace kernel weight and the global regularization parameter after the spatiotemporal constraints. The regularization parameters of the subspace band ridge are substituted into and the band ridge regression coding model is solved to determine the regression weights of each feature subspace, thereby realizing task-state fMRI coding for brain region activity analysis.

2. The task-state fMRI coding method under natural stimulus conditions according to claim 1, characterized in that, Acquire task-oriented fMRI data collected under natural stimulus conditions and construct an fMRI data matrix, specifically including: Based on natural stimulation conditions, the BOLD signals of all voxels at all time points were used as task-state fMRI data. Construct an fMRI data matrix based on the task-state fMRI data.

3. The task-state fMRI coding method under natural stimulus conditions according to claim 1, characterized in that, The objective function of the banded ridge regression coding model is: in, Let v be the regression weight of voxel v in the i-th feature subspace; y v voxels v Time series; X i Let m be the i-th feature subspace; m is the total number of feature subspaces. voxels v The ridge regularization parameter in the i-th feature subspace; ||||2 is the L2 norm.

4. The task-state fMRI encoding method under natural stimulus conditions according to claim 1, characterized in that, Based on the overall candidate set of regularization parameters, the optimal hyperparameters of the banded ridge regression coding model are determined, specifically including: Based on the overall regularization parameter candidate set, multiple sets of candidate kernel weight vectors are set, and a combined kernel is constructed for each candidate kernel weight vector; Based on the combined kernel, N-fold cross-validation is performed on each voxel to search for the optimal global regularization parameter in the candidate set of global regularization parameters; Based on the optimal global regularization parameter corresponding to all candidate kernel weight vectors, compare the overall performance of all candidate kernel weight vectors to determine the optimal subspace kernel weight and the corresponding initial global regularization parameter.

5. The task-state fMRI coding method under natural stimulus conditions according to claim 1, characterized in that, The global regularization parameter after spatiotemporal constraints for: in, u For voxels in the neighborhood of voxel v; N ( v ) represents the spatial neighborhood of voxel v; These are the initial global regularization parameters; For comprehensive weighting; It is a very small constant.

6. A task-oriented fMRI coding system under natural stimulus conditions, characterized in that, The task-oriented fMRI coding system under natural stimulus conditions performs the task-oriented fMRI coding method under natural stimulus conditions according to any one of claims 1-5, wherein the task-oriented fMRI coding system under natural stimulus conditions comprises: The fMRI data matrix construction module is used to acquire task-state fMRI data collected under natural stimulus conditions and construct an fMRI data matrix. The feature subspace construction module is used to construct feature subspaces based on the fMRI data matrix; each feature subspace corresponds to a type of feature; the features include phoneme vector features, phoneme counting features, word counting features, and contextual semantic features of the large language model; A ridge regression coding model construction module is used to construct a ridge regression coding model based on the feature subspace and the fMRI data matrix. The optimal hyperparameter determination module is used to determine the optimal hyperparameters of the banded ridge regression coding model based on the overall regularization parameter candidate set; the optimal hyperparameters include the optimal subspace kernel weights and the corresponding initial global regularization parameters; The spatiotemporal homogeneity constraint module is used to introduce the spatiotemporal homogeneity prior of voxels under natural stimulus conditions, calculate the time series correlation coefficient between each voxel and its neighboring voxels, and determine the global regularization parameter after spatiotemporal constraint based on the initial global regularization parameter. The subspace banded ridge regularization parameter determination module is used to determine the subspace banded ridge regularization parameter based on the optimal subspace kernel weight and the global regularization parameter after the spatiotemporal constraints. The task-oriented fMRI encoding module is used to substitute the regularization parameters of the subspace band ridge into and solve the band ridge regression encoding model, determine the regression weights of each feature subspace, and realize task-oriented fMRI encoding for brain region activity analysis.

7. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the task-oriented fMRI coding method under natural stimulus conditions as described in any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the task-state fMRI coding method under natural stimulus conditions as described in any one of claims 1-5.

9. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the task-state fMRI coding method under natural stimulus conditions as described in any one of claims 1-5.

Images