A brain function sub-region division method, device, equipment and medium

Abstract

The application discloses a brain function subregion division method and device, equipment and medium, and relates to the technical field of biomedical imaging signal processing. The method comprises the following steps: after obtaining initial time characteristics and initial space characteristics of target fMRI data, performing feature extraction on the initial time characteristics and the initial space characteristics by using a trained automatic encoder to obtain deep time characteristics and deep space characteristics, considering spatial, time and hidden nonlinear characteristics, and emphasizing the properties of space-time interaction and complementation; performing spectral clustering analysis on a final similarity matrix which is constructed by iteratively fusing a time similarity matrix and a space similarity matrix of a similarity network, to obtain a brain function subregion division result of a user. The application improves the comprehensiveness and accuracy of brain function subregion division.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of biomedical imaging signal processing, in particular to a brain function sub-region division method, device, equipment and medium. BACKGROUND

[0002] Brain partition refers to a spatially continuous region that is significantly different from its adjacent regions in anatomy or function, including the partition of the whole brain, the whole cerebral cortex and specific brain regions. Brain partition can be divided into anatomy-based partition and function-based partition. Anatomical partition is a region of the brain divided according to its physical structure. This partition is based on the actual anatomical structure of the brain, such as different lobes of the brain (frontal lobe, parietal lobe, occipital lobe, temporal lobe), neural nuclei, etc. Functional partition is a region of consistent function divided according to the activity pattern of the brain. Although anatomical partition provides a clear map of brain structure, it cannot fully demonstrate the activity and interaction of brain regions in function. In addition, anatomical partition may sometimes ignore the plasticity of the brain and the differences between individuals, i.e., the same anatomical structure may perform different functions in different individuals or in different situations. Functional partition can reveal the activity pattern of the brain, thereby providing an in-depth understanding of the function and information processing mode of the brain. Function-based partition divides the brain into functionally homogeneous partitions, and this functional specialization makes it very suitable for exploring brain function and dysfunction. Reliable functional partition can help improve the prediction accuracy of brain diseases.

[0003] Functional magnetic resonance imaging (fMRI) plays a key role in exploring brain functional connectivity. Brain functional connectivity refers to the interaction and coordinated activity between different regions of the brain, which is the basis for understanding how the brain works in an integrated network. Resting-state fMRI (rs-fMRI) allows researchers to explore the intrinsic functional networks of the brain when it is not performing specific tasks. In recent years, brain partition research based on rs-fMRI has developed rapidly. The challenge in this field is the lack of a gold standard for functional partition, so most fMRI-based partition research relies on unsupervised methods.

[0004] In fMRI-based functional partition, traditional unsupervised methods have certain limitations and cannot fully capture the complexity of brain activity, so there is an urgent need for a more comprehensive and accurate brain function sub-region division method, device, equipment and medium. SUMMARY

[0005] The purpose of the present application is to provide a brain function sub-region division method, device, equipment and medium, which can improve the comprehensiveness and accuracy of brain function sub-region division.

[0006] To achieve the above object, the application provides the following scheme.

[0007] In a first aspect, the application provides a brain function subregion division method, comprising:

[0008] obtaining original fMRI data of a user;

[0009] preprocessing the original fMRI data to obtain target fMRI data;

[0010] extracting fMRI signals corresponding to each voxel in a region to be divided in the target fMRI data to obtain initial time features of the target fMRI data; extracting fMRI signals corresponding to each brain region in the target fMRI data to obtain a brain region time signal matrix; and determining initial spatial features of the target fMRI data according to the initial time features and the brain region time signal matrix;

[0011] using a trained autoencoder to perform feature extraction on the initial time features and the initial spatial features of the target fMRI data respectively to obtain deep time features and deep spatial features;

[0012] constructing a similarity network; the vertices of the similarity network correspond one-to-one to the voxels in the region to be divided; the weights of the edges of the similarity network are determined by a time similarity matrix and a spatial similarity matrix; the time similarity matrix is determined by the Pearson correlation coefficients of the deep time features of two voxels connected by an edge; and the spatial similarity matrix is determined by the Pearson correlation coefficients of the deep spatial features of two voxels connected by an edge;

[0013] iteratively fusing and calculating a final similarity matrix according to the time similarity matrix and the spatial similarity matrix;

[0014] for a user, using an NCut spectral clustering algorithm to analyze the final similarity matrix to obtain an individual function subregion division result of the user; the individual function subregion division result comprises a plurality of function subregions.

[0015] In a second aspect, the application provides a brain function subregion division device, comprising the following modules.

[0016] an original fMRI data obtaining module, configured to obtain original fMRI data of a user;

[0017] a preprocessing module, configured to preprocess the original fMRI data to obtain target fMRI data;

[0018] An initial feature extraction module is configured to: extract fMRI signals corresponding to each voxel in a region to be divided in target fMRI data to obtain initial temporal features of the target fMRI data; extract fMRI signals corresponding to each brain region in the target fMRI data to obtain a brain region temporal signal matrix; and determine initial spatial features of the target fMRI data according to the initial temporal features and the brain region temporal signal matrix.

[0019] A deep feature extraction module is configured to: extract deep temporal features and deep spatial features from the initial temporal features and the initial spatial features of the target fMRI data by using a trained auto-encoder.

[0020] A similarity network construction module is configured to: construct a similarity network; vertices of the similarity network correspond to voxels in the region to be divided one by one; weights of edges of the similarity network are determined by a temporal similarity matrix and a spatial similarity matrix; the temporal similarity matrix is determined by a Pearson correlation coefficient of deep temporal features of two voxels connected by an edge; and the spatial similarity matrix is determined by a Pearson correlation coefficient of deep spatial features of two voxels connected by an edge.

[0021] A final similarity matrix calculation module is configured to: iteratively fuse and calculate a final similarity matrix according to the temporal similarity matrix and the spatial similarity matrix.

[0022] A functional subregion division module is configured to: for a user, analyze the final similarity matrix by using an NCut spectral clustering algorithm to obtain an individual functional subregion division result of the user; and the individual functional subregion division result includes a plurality of functional subregions.

[0023] In a third aspect, the present application provides a computer device, comprising: 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 brain functional subregion division method.

[0024] In a fourth aspect, the present application provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executable on a processor to implement the above brain functional subregion division method.

[0025] According to the embodiments provided in the present application, the following technical effects are disclosed:

[0026] The application provides a brain function sub-region division method, device, equipment and medium, which is realized based on an end-to-end joint optimization framework, after initial time characteristics and initial space characteristics of target fMRI data are obtained, the initial time characteristics and the initial space characteristics are subjected to feature extraction by using a trained automatic encoder to obtain deep time characteristics and deep space characteristics, the spatial, temporal and hidden nonlinear characteristics (the deep time characteristics and the deep space characteristics) are considered, and the space-time interaction and complementary properties are emphasized; the final similarity matrix of the time similarity matrix and the space similarity matrix of the similarity network is subjected to spectral clustering analysis, and an individual function sub-region division result of a user is obtained, the spatial and temporal characteristics of brain activity are fused into deep embedding clustering, the deep embedding clustering is composed of a stacked automatic encoder and spectral clustering, the contribution of the spatial and temporal characteristics can be adaptively adjusted, the single nature of a traditional method depending on a single feature is solved, and the comprehensiveness and accuracy of brain function sub-region division are improved. BRIEF DESCRIPTION OF DRAWINGS

[0027] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0028] Figure 1 The application environment diagram of a brain function sub-region division method in an embodiment of the present application;

[0029] Figure 2 The flowchart of a brain function sub-region division method provided in an embodiment of the present application;

[0030] Figure 3 The repeatability comparison result diagram of different methods in all possible partition numbers in an ablation experiment provided in an embodiment of the present application; Figure 3 (a), (b) and (c) in the figure are respectively the repeatability comparison results of Dice coefficients, intraclass correlation coefficients and normalized mutual information of different methods in all possible partition numbers under a repeatability comparison strategy between two scans of the same user; Figure 3 (d), (e) and (f) are respectively the repeatability comparison results of Dice coefficients, intraclass correlation coefficients and normalized mutual information of different methods in all possible partition numbers under a repeatability comparison strategy between two group-level results; Figure 3 (g), (h) and (i) are respectively the repeatability comparison results of Dice coefficients, intraclass correlation coefficients and normalized mutual information of different methods in all possible partition numbers under a repeatability comparison strategy between paired individuals; Figure 3(j), (k), (l) in FIG. 6 are respectively the repeatability comparison results of the Dice coefficient, the intraclass correlation coefficient and the normalized mutual information of different methods under the leave-one-out strategy in all possible partition numbers;

[0031] Figure 4 FIG. 7 is a schematic diagram of the functional homogeneity comparison results of different methods in the ablation experiment provided by an embodiment of the present application;

[0032] Figure 5 FIG. 8 is a schematic diagram of the difference between the graph theory indexes of healthy users and Parkinson's disease users provided by an embodiment of the present application; Figure 5 (a) in FIG. 8 is the difference between the modularity indexes of healthy users and Parkinson's disease users; Figure 5 (b) in FIG. 8 is the difference between the small-world indexes of healthy users and Parkinson's disease users;

[0033] Figure 6 FIG. 9 is a schematic diagram of the relationship between the changes of the graph theory indexes and the unified Parkinson's disease rating scale provided by an embodiment of the present application; Figure 6 (a) in FIG. 9 is the relationship between the changes of the modularity indexes and the unified Parkinson's disease rating scale; Figure 6 (b) in FIG. 9 is the relationship between the changes of the small-world indexes and the unified Parkinson's disease rating scale;

[0034] Figure 7 FIG. 10 is a schematic diagram of the functional modules of a brain function subregion division device provided by an embodiment of the present application;

[0035] Figure 8 FIG. 11 is a schematic diagram of the structure of a computer device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0036] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0037] The above purposes, features and advantages of the present application can be more obvious and easy to understand. The present application will be described in further detail below with reference to the drawings and specific embodiments.

[0038] Unsupervised methods are diverse and rich, including various clustering techniques such as hierarchical clustering, K-means clustering and spectral clustering, as well as region growing, boundary mapping and decomposition techniques such as independent component analysis. In recent years, deep neural networks (DNN) have also made progress in brain parcellation. Although DNN has become the preferred technology for automatic brain segmentation and brain disease prediction based on structural MRI using anatomical atlas / mask or disease state label in supervised methods, its potential in fMRI-based parcellation remains to be further explored. Existing research includes supervised brain parcellation using deep residual network to predict the probability of a voxel belonging to a functional network pre-determined by ICA, demonstrating the practicality of deep learning algorithms. However, research based on unsupervised deep learning methods is relatively less.

[0039] Input features can generally be divided into two categories: time-based features and space-based features. Time-based feature division relies on blood oxygen level dependent (BOLD) amplitude (e.g., time series of voxels or its transformed variants, activation values of voxels) to characterize the local functional activity of the region to be parcellated (i.e., target region). On the other hand, space-based feature division captures the long-range connections between the target region and other regions (i.e., reference regions) to outline the global interaction patterns of the target region.

[0040] In fMRI-based functional parcellation, traditional unsupervised methods usually only focus on time features or space features. This includes clustering analysis using correlation or similarity of time series, or defining brain regions based on spatial pattern distribution. However, these single-dimensional methods have certain limitations, such as the inability to fully capture the complexity of brain activity. Therefore, it is particularly important to consider a comprehensive approach that simultaneously integrates spatial and temporal features. This comprehensive analysis strategy is expected to provide a more comprehensive and accurate mapping of brain functional regions, providing richer insights into the understanding of neurodegenerative diseases and other brain dysfunction.

[0041] Therefore, the present application integrates local temporal characteristics, global spatial characteristics and hidden nonlinear characteristics from three perspectives, and considers the spatio-temporal interaction to perform brain functional sub-region division. The framework opens up new possibilities for functional sub-region division of brain regions through deep learning and adaptive spatio-temporal feature fusion technology, and provides a new tool for brain science research.

[0042] The brain functional sub-region division method provided by the embodiments of the present application can be applied to, for example Figure 1The application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data required by the server 104 to process. The data storage system can be set up separately, or integrated on the server 104, or placed on the cloud or other servers. The terminal 102 can send the user's original fMRI data to the server 104, and after the server 104 receives the user's original fMRI data, the server 104 pre-processes the original fMRI data of the user to obtain target fMRI data, and the target fMRI data is extracted to obtain initial time features and initial space features. The initial time features and the initial space features are extracted by using the trained automatic encoder to obtain deep time features and deep space features, a similarity network is constructed, the final similarity matrix is iteratively calculated according to the time similarity matrix and the space similarity matrix in the similarity network, and the NCut spectral clustering algorithm is used to analyze the final similarity matrix to obtain the individual functional subregion division result of the user. The server 104 can feed back the individual functional subregion division result obtained for the user to the terminal 102. In addition, in some embodiments, the brain functional subregion division method can also be implemented by the server 104 or the terminal 102 alone, such as directly performing functional subregion division on the user's original fMRI data by the terminal 102, or obtaining the user's original fMRI data from the data storage system by the server 104 and performing functional subregion division on the user's original fMRI data.

[0043] Among them, the terminal 102 can be but not limited to various desktop computers, notebook computers, smart phones and tablet computers, Internet of Things devices. The server 104 can be implemented by an independent server or a server cluster composed of multiple servers, and can also be a cloud server.

[0044] In an exemplary embodiment, as Figure 2 shown, a brain functional subregion division method is provided, which is executed by a computer device, specifically by a terminal or a server, etc. Computer device alone, or by a terminal and a server together, in the embodiment of the present application, take the server 104 in Figure 1 as an example to illustrate the method, including the following steps 201 to 207. Among them:

[0045] Step 201: obtaining the user's original fMRI data.

[0046] Step 202: pre-processing the original fMRI data to obtain target fMRI data.

[0047] Step 203: Extracting the fMRI signal corresponding to each voxel in the target fMRI data to be divided region to obtain the initial time feature of the target fMRI data; extracting the fMRI signal corresponding to each brain region in the target fMRI data to obtain the brain region time signal matrix; determining the initial spatial feature of the target fMRI data according to the initial time feature and the brain region time signal matrix.

[0048] Step 204: Using the trained auto-encoder to extract features from the initial time feature and the initial spatial feature of the target fMRI data respectively to obtain deep time features and deep spatial features.

[0049] Step 205: Constructing a similarity network; the vertices of the similarity network correspond one-to-one to the voxels in the region to be divided; the weight of the edge of the similarity network is determined by the time similarity matrix and the spatial similarity matrix; the time similarity matrix is determined by the Pearson correlation coefficient of the deep time features of the two voxels connected by the edge; the spatial similarity matrix is determined by the Pearson correlation coefficient of the deep spatial features of the two voxels connected by the edge.

[0050] Step 206: Iteratively fusing the final similarity matrix according to the time similarity matrix and the spatial similarity matrix.

[0051] Step 207: For a user, using the NCut spectral clustering algorithm to analyze the final similarity matrix to obtain the individual functional subregion division result of the user; the individual functional subregion division result includes several functional subregions.

[0052] Implementing the above steps 201 to 207, considering the spatial, temporal and hidden nonlinear features (deep time features and deep spatial features), emphasizing the spatiotemporal interaction and complementary properties, the present application fuses the spatial and temporal features of brain activity into deep embedding clustering, which is composed of stacked auto-encoder and spectral clustering, can adaptively adjust the contribution of spatial and temporal features, solves the single nature of traditional methods relying on single feature, improves the comprehensiveness and accuracy of brain functional subregion division. In addition, the present application can also determine the group-level functional subregion division result according to the individual functional subregion division result of different users, which has the flexibility of partitioning at individual and group levels.

[0053] In step 201 of the present application, a set of original fMRI data of 176 healthy users is obtained, and for each user's original fMRI data, preprocessing is required, then the preprocessing process in step 202 is as follows:

[0054] (1) Spatial artifact removal, which removes the artifacts caused by magnetic field inhomogeneity from the original fMRI images through techniques such as distortion correction. (2) Time slice correction, which is used to solve the problem of time difference between slices in the same fMRI data during scanning. (3) Head motion correction, which eliminates the impact of head movement on the acquisition of magnetic resonance data, as the user remains still, but there is still slight movement of the head due to psychological factors such as heartbeat or breathing. (4) Registration, which aligns the user's fMRI data to its structural imaging, facilitating subsequent analysis. (5) Spatial standardization, which reduces the differences between users by aligning the brain space of each user to a standard reference space, ensuring data consistency and comparability during group-level analysis. (6) Linear drift reduction, which reduces the noise interference caused by various sources of other signals. (7) Finally, spatial smoothing is performed to reduce noise and random variation in the fMRI signal.

[0055] In another exemplary embodiment of the present application, in step 203, the fMRI signal corresponding to each brain region in the target fMRI data is extracted to obtain a brain region time signal matrix, which can specifically include steps 301-303 as follows:

[0056] Step 301: Extract the fMRI signal corresponding to each brain region in the cortical region of the target fMRI data using the surface-based multi-modal brain region atlas to obtain the fMRI signal time point matrix corresponding to each brain region in the cortical region.

[0057] Step 302: Extract the fMRI signal corresponding to each brain region in the subcortical region of the target fMRI data using the Desikan-Killiany atlas template to obtain the fMRI signal time point matrix corresponding to each brain region in the subcortical region.

[0058] Step 303: Determine the brain region time signal matrix according to the fMRI signal time point matrix corresponding to each brain region in the cortical region and the fMRI signal time point matrix corresponding to each brain region in the subcortical region.

[0059] In step 203, the initial spatial feature of the target fMRI data is determined according to the initial temporal feature and the brain region time signal matrix, which specifically includes: calculating the Pearson correlation coefficient according to the initial temporal feature and the brain region time signal matrix; and taking the Pearson correlation coefficient as the initial spatial feature of the target fMRI data.

[0060] The initial time feature refers to the number of voxels of each user's to-be-partitioned region multiplied by the time signal matrix of the time point. The time feature mainly considers the local information of the to-be-partitioned region, that is, the region in the brain. The initial space feature refers to first calculating the Pearson correlation coefficient of the to-be-partitioned region time signal matrix of each user's to-be-partitioned region voxels multiplied by the time point and the signal matrix of the brain region multiplied by the time point, and thus obtaining the connection feature between the to-be-partitioned region and other regions of the brain, which is the space feature. The initial space feature considers the global information.

[0061] The brainstem is taken as an example to perform functional sub-region division. The existing data of the brainstem part are all voxel-level data. The fMRI data in the HCP data set is used, and a 3472 (number of brainstem voxels) by 900 (time points) signal matrix is obtained for each user. The 3472 by 900 signal matrix is taken as the initial time feature. The amount of data of other regions of the brain in the voxel level is too large, and needs to be reduced. Since the HCP data set has been preprocessed to generate a surface, the HCP-MMP (HCP-Multimodal Parcellation) based on the surface is used to extract the ROI (region of interest) signal of the cortical region. The atlas contains 360 multimodal brain cortical regions, and the fMRI signal time point matrix corresponding to each brain region in the cortical region is extracted. The Desikan-Killiany atlas template is used for the subcortical part. The template divides the subcortical region into 18 brain region areas, and the fMRI signal time point matrix corresponding to each brain region in the subcortical region is extracted. Using the two templates, a 378 (ROI) by 900 (time points) signal matrix can be obtained for each user. The 378 (ROI) by 900 (time points) signal matrix is taken as the brain region time signal matrix, which is used as the data for generating the functional sub-region subsequently.

[0062] After obtaining the initial time feature and the brain region time signal matrix, the Pearson correlation coefficient of the two is calculated as the initial space feature of the target fMRI data.

[0063] The end-to-end joint optimization framework (deep learning framework) integrates a novel adaptive similarity network fusion module, which fuses the spatial and temporal features of brain activity into deep embedding clustering, which is composed of a stacked autoencoder and spectral clustering.

[0064] In another exemplary embodiment of the present application, the autoencoder needs to be trained before step 204 to obtain a trained autoencoder.

[0065] The autoencoder is composed of an encoder and a decoder. The encoder maps the input initial time feature or initial space feature of the to-be-partitioned region to a vector x through a nonlinear transformation, and maps it to a hidden representation h:

[0066] h = f(W1x + b1) (1);

[0067] where f is a nonlinear activation function, W1represents the encoding weight matrix and b1is the encoding bias. The decoder reconstructs the input data from the hidden representation h, aiming to minimize the reconstruction error:

[0068] x' = f(W2h + b2) (2);

[0069] where x' is the reconstructed input, i.e., the output of the decoder; W2and b2are the decoding weight and bias, respectively. The training process involves minimizing the loss function Mean Squared Error (MSE). If the input vector x is composed of N features, the loss function of the autoencoder is:

[0070]

[0071] where L(x, x') represents the loss function, N represents all elements of the matrix in the initial temporal features or initial spatial features of the region to be divided, x i represents the i-th element of the matrix in the initial temporal features or initial spatial features, x i ' represents the reconstructed input corresponding to the i-th element of the matrix in the initial temporal features or initial spatial features.

[0072] The similarity network is represented as a graph G = (V, E). The vertices V correspond to the voxels {x1, x2,... x n} in the region to be divided, n represents the number of voxels in the region to be divided, and the edges E are weighted according to the similarity between the voxels, and the weight is determined by an n x n similarity matrix W, W(i, j) represents the similarity between voxels x i and x j , the similarity matrix W is obtained based on the Pearson correlation coefficient of the depth temporal features and the depth spatial features, respectively, and it is set that the voxels are not adjacent and the similarity is zero. In this embodiment, the similarity matrix obtained based on the Pearson correlation coefficient of the depth temporal features is named as the temporal similarity matrix, and the similarity matrix obtained based on the Pearson correlation coefficient of the depth spatial features is named as the spatial similarity matrix.

[0073] In another exemplary embodiment of the present application, the above step 206 is replaced by steps 401 to 403 as follows:

[0074] Step 401: normalizing the temporal similarity matrix and the spatial similarity matrix respectively to obtain a temporal normalized matrix and a spatial normalized matrix;

[0075] Step 402: calculating the time neighboring voxel similarity matrix and the space neighboring voxel similarity matrix according to the time similarity matrix and the space similarity matrix respectively;

[0076] Step 403: performing iterative fusion according to the time neighboring voxel similarity matrix, the space neighboring voxel similarity matrix, the time normalization matrix and the space normalization matrix to obtain a final similarity matrix.

[0077] The time similarity matrix and the space similarity matrix need to be further normalized. One method of normalization is shown in the following formula:

[0078]

[0079] wherein P(i, j) represents a normalized matrix element value between voxels x i and x j , W(i, k) represents similarity between voxels x i and x k , and k is the serial number of neighboring voxel x i in the neighboring voxel set of voxel x k .

[0080] When processing graph G, the method uses K-Nearest Neighbor algorithm to quantify the similarity matrix to obtain the time neighboring voxel similarity matrix and the space neighboring voxel similarity matrix, as shown in the following formula:

[0081]

[0082] wherein S(i, j) represents an element value in the neighboring voxel similarity matrix between voxels x i and x j , N i is the neighboring voxel set of voxel x i , and the neighboring voxel of each voxel is determined by K-Nearest Neighbor algorithm.

[0083] The core assumption of K-Nearest Neighbor algorithm is that, for a given data point, it has higher similarity with nearby data points and less reliable similarity with distant data points. In the key iterative step in the fusion method, the fusion matrix is constantly updated by the following formula to reflect the similarity between data points:

[0084]

[0085] wherein represents the fusion matrix after t iterations of fusion of the first type of data feature, i.e. the time feature fusion matrix; represents the fusion matrix after t iterations of fusion of the second type of data feature, i.e. the space feature fusion matrix; denotes the spatial feature fusion matrix obtained after t-1 iterations, denotes the temporal feature fusion matrix obtained after t-1 iterations; S (1) denotes the temporal neighboring voxel similarity matrix, S (2) denotes the spatial neighboring voxel similarity matrix; (S (1) ) T and (S (2) ) T denote the transpose of the temporal neighboring voxel similarity matrix and the spatial neighboring voxel similarity matrix, respectively.

[0086] In the iteration process, the two matrices (temporal feature fusion matrix and spatial feature fusion matrix) will influence and fuse with each other in the fusion process. After n steps, the final similarity matrix P (g) is calculated as follows:

[0087]

[0088] In another exemplary embodiment of the present application, in order to properly consider the interaction between the temporal feature and the spatial feature, and accurately reflect the role of the two features in data analysis and processing, the respective contribution of the spatio-temporal feature needs to be quantified in this paper, not only to determine which feature is more important, but also to more deeply understand how the temporal and spatial features interact with each other and demonstrate their unique influence in different application scenarios. To this end, an adaptive learning mechanism is introduced, that is, by adjusting a parameter a to balance the importance and contribution between the temporal feature and the spatial feature .

[0089]

[0090] where P(c) denotes the final similarity matrix. Formula (9) describes an adaptive learning framework, where a is a parameter between 0 and 1, representing the relative importance of the temporal feature and the spatial feature. When a is close to 1, the model is more biased towards the temporal feature; when a is close to 0, it is more biased towards the spatial feature. This dynamic adjustment mechanism enables the model to flexibly adjust the importance of spatio-temporal features according to different data sets and application requirements.

[0091] In another exemplary embodiment of the present application, the above step 207 includes the following steps 501-505:

[0092] Step 501: absolute value conversion is performed on the final similarity matrix to construct an adjacency matrix and a degree matrix;

[0093] Step 502: a Laplacian matrix is calculated;

[0094] Step 503: For each voxel, a normalized Laplacian matrix corresponding to each voxel is constructed according to the degree matrix and the Laplacian matrix;

[0095] Step 504: For each voxel, the eigenvalue of the normalized Laplacian matrix corresponding to the voxel is calculated, and all eigenvalues are arranged in ascending order, and the eigenvectors corresponding to the first k eigenvalues are taken to construct an eigenmatrix; k is less than the number of voxels; the eigenmatrix is an n x k matrix; n represents the number of voxels;

[0096] Step 505: The eigenmatrix is subjected to NCut spectral clustering to obtain the individual functional subregion division result of the user, specifically including: using a K-means clustering algorithm, the eigenmatrix is subjected to NCut spectral clustering to obtain the individual functional subregion division result of the user.

[0097] Based on the above similarity matrix generation method, for a given input sample set D = {x1, x2,... x n}a final similarity matrix P is constructed, wherein x n represents the nth voxel in the region to be divided.

[0098] The absolute value of the final similarity matrix P is converted, the similarity value of the non-adjacent voxel is set to zero, and the similarity value of the adjacent voxel is retained, and then the adjacency matrix W and the degree matrix D are obtained.

[0099] When the Laplacian matrix L is calculated, a normalized Laplacian matrix is constructed, and the calculation formula is as follows:

[0100]

[0101] Subsequently, the eigenvalues of the normalized Laplacian matrix L n are calculated and arranged in ascending order. The eigenvectors f corresponding to the first k eigenvalues are selected to form an eigenmatrix. These eigenvectors f are then normalized to generate an n x k eigenmatrix H. In the eigenmatrix H, each row represents a k-dimensional sample, and there are n samples. Finally, a clustering method is used to cluster the k-dimensional samples to obtain a set of cluster partitions C = {c1, c2,... c n}, effectively distributing the samples in the data set D into k different clusters to achieve data clustering. wherein c n represents the functional subregion corresponding to the nth voxel in the region to be divided. Preferably, the present embodiment can use a K-means clustering algorithm to cluster the k-dimensional samples. The cut value of the NCut spectral clustering method is considered as a key indicator in graph theory, which can effectively evaluate the quality and performance of graph partitioning.

[0102] For the adaptive similarity network fusion module, the application proposes a new loss function which comprehensively considers the reconstruction loss value and the NCut cut value of the spatio-temporal features, and is used to guide the training process of the deep learning framework. The loss function formula is as follows:

[0103]

[0104] Wherein, L represents the loss function of the adaptive similarity network fusion module, x1 and are the initial time features of the input and the reconstructed region to be divided, x2 and are the initial space features of the input and the reconstructed region to be divided; F is the membership matrix of the voxels in the region to be divided, each row of the membership matrix represents the functional subregion division result corresponding to the voxel, the membership matrix has q columns, q is the number of functional subregions in the individual functional subregion division result, when the voxel belongs to the functional subregion, the corresponding element value in the matrix is 1, and when the voxel does not belong to the functional subregion, the corresponding element value in the matrix is 0, so the membership matrix is an n x q matrix; D and L are the degree matrix and the normalized Laplacian matrix of the similarity matrix after deep spatio-temporal fusion, and tr() represents the trace of the matrix. The loss function provides a new perspective for the brain function division field, and can comprehensively and carefully analyze various characteristics of the brain.

[0105] In another exemplary embodiment of the application, after obtaining the individual functional subregion division result of the user, the brain functional subregion division method further comprises steps 601 to step 604:

[0106] Step 601: obtaining individual functional subregion division results of different users;

[0107] Step 602: for each of the individual functional subregion division results, constructing an adjacency matrix corresponding to the individual functional subregion division result according to the region division result corresponding to each voxel in the individual functional subregion division result;

[0108] Step 603: constructing a group-level graph structure according to the adjacency matrices corresponding to all the individual functional subregion division results; the vertices of the group-level graph structure correspond one-to-one to the voxels in the region to be divided; the weight of each edge of the group-level graph structure is determined by the same allocation frequency of the two vertices connected by the edge; the same allocation frequency represents the number of voxels corresponding to the two vertices being divided into the same functional subregion in all individual functional subregion division results;

[0109] Step 604: performing clustering analysis on the group-level graph structure to obtain a group-level functional subregion division result.

[0110] To obtain the individual-level and group-level partition results, the present embodiment adopts a two-level analysis method, aiming to explore the function and structure of the brainstem. In the first-level analysis, the focus is on the brainstem structure of each individual, and a N x N adjacency matrix (where N represents the number of voxels in the region to be partitioned) is constructed to identify functionally consistent regions. This process involves analyzing the data of the region to be partitioned for each user to extract the similarity between vertices. In the second-level analysis, the focus shifts to the group level, and a stability graph (group-level graph structure) is constructed by integrating the partition results of all individuals. In this group-level graph structure, the weight of the edge is determined based on the frequency of the allocation of the same block in all users i and v j . Finally, a clustering algorithm is applied to subdivide the group-level graph structure, resulting in the group-level block partition and the group-level functional subregion partition result (group-level brainstem functional subregion template).

[0111] To verify the performance of the brain functional subregion partition method proposed in the present application, two indicators, repeatability and functional homogeneity, are used. The repeatability indicators include Dice coefficient, intra-class correlation coefficient (ICC), and normalized mutual information (NMI). The comparison process mainly involves four different comparison strategies, namely the repeatability between two scans of the same user, the repeatability between two group-level results, the repeatability between pairs of individuals, and the repeatability between individuals using the leave-one-out method. Ablation experiments are conducted using different methods, including deep learning spatiotemporal feature fusion embedding (the brain functional subregion partition method proposed in the present application), deep learning spatiotemporal feature importance embedding, deep temporal feature (deep learning framework only uses temporal features), deep spatial feature (deep learning framework only uses spatial features), spatiotemporal feature fusion embedding (only uses temporal and spatial features without deep learning framework), temporal feature (only uses temporal features without deep learning framework), and spatial feature (only uses spatial features without deep learning framework), as shown in Figure 3 , the repeatability provided by the present application is the highest in the ablation experiment. The method using the deep learning framework outperforms the methods not using deep learning technology. The method fusing spatiotemporal features can produce more excellent results than the methods using only spatial or temporal features. This finding emphasizes the importance of deep features, temporal features, and spatial features in brain region function partitioning, indicating that the comprehensive use of the three is crucial for improving analysis accuracy. As shown in Figure 4 , the functional homogeneity provided by the present application is the highest in the ablation experiment.

[0112] By fusing temporal and spatial features, it is possible to better extract and integrate information in different dimensions, thereby improving functional homogeneity.

[0113] In order to study the practicality of the functional area division method, an in-depth analysis of the brain networks of Parkinson's disease (PD) users and healthy users was conducted based on the proposed functional map. Nine graph theory indicators in four categories were selected: modularity, clustering coefficient, local efficiency, transitivity, global efficiency, betweenness centrality, eigenvector centrality, synchrony, and small-world index. By calculating and comparing these indicators, a two-sample t-test was used to statistically analyze the brain networks of PD patients and healthy users. Figure 5 As shown in Figure 2, the modularity and small-world indicators of healthy users are significantly higher than those of PD patients. Figure 6 As shown, modularity and small-worldness indices showed significant negative correlations with the Unified Parkinson's Disease Rating Scale (UPDRS). The negative correlation between modularity and UPDRS (p = 0.0060, r = -0.6497) suggests that as PD symptoms worsen, the modularity of brain networks may gradually decrease. This may be because the progression of PD leads to a more complex brain network structure, blurring the boundaries between previously functionally independent modules, thereby affecting the effectiveness and functional specificity of brain networks. The negative correlation between small-worldness and UPDRS (p = 0.0485, r = -0.4612) suggests that as PD symptoms worsen, brain networks may experience problems balancing local clustering and global information transfer, leading to decreased network efficiency.

[0114] The performance of this method was further studied, and the classification prediction of PD was made based on the functional template proposed by the functional area segmentation method. Regardless of the machine learning method used, the functional segmentation-based method proposed in this application consistently outperformed the anatomical template-based segmentation in terms of accuracy. In K-nearest neighbor, the functional template was 0.84 and the anatomical template was 0.79; in discriminant analysis, the functional template was 0.87 and the anatomical template was 0.74; in logistic regression, the accuracy of the functional template was 0.78 and the anatomical template was 0.65. In naive Bayes, the functional template was 0.82 and the anatomical template was 0.72; in random forest, the functional template was 0.83 and the anatomical template was 0.76. In the context of Parkinson's disease, the functional activity of certain brain regions may have changed in the early stages of the disease, even if there are no obvious abnormalities in their structure. By capturing these early functional changes, functional maps can help diagnose and understand the development of the disease earlier.

[0115] The present application aims to divide the brain into sub-regions based on fMRI signals according to functional characteristics, so as to explore the basic organization and function of the brain. The brain functional sub-region division method proposed in the present application is a deep spatiotemporal embedding division method, which is realized based on an end-to-end joint optimization framework. The end-to-end joint optimization framework takes into account the spatial, temporal and hidden nonlinear characteristics at the same time, and emphasizes the spatiotemporal interaction and complementary nature. The end-to-end joint optimization framework integrates a novel adaptive similarity network fusion module, which fuses the spatial and temporal characteristics of brain activity into deep embedding clustering, which is composed of a stacked autoencoder and spectral clustering, and can adaptively adjust the contribution of spatial and temporal characteristics, solving the single nature of traditional methods which rely on a single feature. The present application also has flexibility in individual and group level partitioning. The present application provides a new method for brain functional region division, provides an important basis for brain function research, and can support clinical application research of brain diseases.

[0116] The present application also provides an application scenario applying the above-mentioned brain functional sub-region division method. Specifically: the brain functional sub-region division method provided in the present embodiment can be applied in a brainstem functional partitioning scenario. The brainstem functional partitioning scenario includes a data acquisition link, a functional sub-region division link and a group level division link; the original fMRI data of a user enters the functional sub-region division link from the data acquisition link, obtains the corresponding individual functional sub-region division result, and enters the downstream group level division link to determine the group level functional sub-region division result according to the individual functional sub-region division results of different users. The brain functional sub-region division method provided in the present embodiment belongs to the functional sub-region division link. Specifically, in the process of the functional sub-region division link for the original fMRI data of a user, the original fMRI data can be preprocessed to obtain target fMRI data, the target fMRI data is subjected to feature extraction to obtain initial temporal features and initial spatial features, the trained autoencoder is used to extract features from the initial temporal features and the initial spatial features respectively to obtain deep temporal features and deep spatial features, a similarity network is constructed, the final similarity matrix is calculated according to the temporal similarity matrix and the spatial similarity matrix in the similarity network, and the NCut spectral clustering algorithm is used to analyze the final similarity matrix to obtain the individual functional sub-region division result of the user.

[0117] Based on the same inventive concept, the present embodiment also provides a brain functional sub-region division device for implementing the above-mentioned brain functional sub-region division method. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme described in the above-mentioned method, and therefore the specific limitations in one or more brain functional sub-region division device embodiments provided below can refer to the limitations of the brain functional sub-region division method described above, which will not be repeated here.

[0118] In an exemplary embodiment, as shown in Figure 7 A brain function sub-region division apparatus is provided and includes the following modules.

[0119] An original fMRI data acquisition module T1 is configured to acquire original fMRI data of a user.

[0120] A preprocessing module T2 is configured to preprocess the original fMRI data to obtain target fMRI data.

[0121] An initial feature extraction module T3 is configured to extract fMRI signals corresponding to each voxel in a region to be divided in the target fMRI data to obtain initial temporal features of the target fMRI data, extract fMRI signals corresponding to each brain region in the target fMRI data to obtain a brain region temporal signal matrix, and determine initial spatial features of the target fMRI data according to the initial temporal features and the brain region temporal signal matrix.

[0122] A deep feature extraction module T4 is configured to extract features from the initial temporal features and the initial spatial features of the target fMRI data using a trained autoencoder to obtain deep temporal features and deep spatial features.

[0123] A similarity network construction module T5 is configured to construct a similarity network, wherein vertices of the similarity network correspond one-to-one to voxels in the region to be divided, weights of edges of the similarity network are determined by a temporal similarity matrix and a spatial similarity matrix, the temporal similarity matrix is determined by a Pearson correlation coefficient of deep temporal features of two voxels connected by an edge, and the spatial similarity matrix is determined by a Pearson correlation coefficient of deep spatial features of two voxels connected by an edge.

[0124] A final similarity matrix calculation module T6 is configured to iteratively fuse and calculate a final similarity matrix according to the temporal similarity matrix and the spatial similarity matrix.

[0125] A function sub-region division module T7 is configured to, for a user, analyze the final similarity matrix using an NCut spectral clustering algorithm to obtain an individual function sub-region division result of the user, and the individual function sub-region division result includes a plurality of function sub-regions.

[0126] In an exemplary embodiment, a computer device, which can be a server or a terminal, is provided, and an internal structure diagram of the computer device can be as shown in Figure 8As shown in the figure. The computer device includes a processor, a memory, an input / output interface (I / O for short) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capability. The memory of the computer device includes a non-volatile storage medium and an 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 the computer program in the non-volatile storage medium. The database of the computer device is used to store functional subregion division data. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals through network connection. The computer program is executed by the processor to realize a brain functional subregion division method.

[0127] Those skilled in the art can understand that, Figure 8 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0128] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to realize the steps in each of the above method embodiments.

[0129] In one exemplary embodiment, a computer readable storage medium is provided, storing a computer program, which is executed by a processor to realize the steps in each of the above method embodiments.

[0130] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant regulations.

[0131] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when the computer program is executed, the processes of the above-mentioned embodiments of the methods can be included. Any reference to memory, database or other medium used in the embodiments provided in the present 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 storage, high-density embedded non-volatile memory, resistive memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration but not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0132] The database involved in the embodiments provided in the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a blockchain, etc., without being limited thereto. The processor involved in the embodiments provided in the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., without being limited thereto.

[0133] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, all possible combinations of the technical features in the above embodiments are not described, however, as long as the combinations of the technical features do not exist contradictory, they should be considered as the scope of the present application.

[0134] The principles and implementation modes of the present application are described by applying specific examples herein, and the above-mentioned embodiments are only used to help understand the method and its core idea of the present application; meanwhile, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range can be changed. In conclusion, the content of the present application should not be understood as a limitation.

Claims

1. A method of dividing a brain function subregion, characterized by, The brain function subregion division method comprises: obtaining original fMRI data of a user; preprocessing the original fMRI data to obtain target fMRI data; extracting fMRI signals corresponding to each voxel in a region to be divided in the target fMRI data to obtain initial temporal features of the target fMRI data; extracting fMRI signals corresponding to each brain region in the target fMRI data to obtain a brain region temporal signal matrix; and determining initial spatial features of the target fMRI data according to the initial temporal features and the brain region temporal signal matrix; extracting features from the initial temporal features and the initial spatial features of the target fMRI data by using a trained automatic encoder to obtain deep temporal features and deep spatial features; constructing a similarity network; the vertices of the similarity network correspond one-to-one to voxels in the region to be divided; the weights of the edges of the similarity network are determined by a temporal similarity matrix and a spatial similarity matrix; the temporal similarity matrix is determined by the Pearson correlation coefficients of the deep temporal features of two voxels connected by an edge; and the spatial similarity matrix is determined by the Pearson correlation coefficients of the deep spatial features of two voxels connected by an edge; iteratively fusing the temporal similarity matrix and the spatial similarity matrix to calculate a final similarity matrix; for a user, analyzing the final similarity matrix by using an NCut spectral clustering algorithm to obtain an individual function subregion division result of the user; the individual function subregion division result comprises a plurality of function subregions.

2. The method of brain functional subsegmentation according to claim 1, wherein, After the step of obtaining the individual function subregion division result of the user, the brain function subregion division method further comprises: obtaining individual function subregion division results of different users; for each individual function subregion division result, constructing an adjacency matrix corresponding to the individual function subregion division result according to the region division result corresponding to each voxel in the individual function subregion division result; constructing a group-level graph structure according to the adjacency matrices corresponding to all the individual function subregion division results; the vertices of the group-level graph structure correspond one-to-one to voxels in the region to be divided; the weight of each edge of the group-level graph structure is determined by the same allocation frequency of two vertices connected by the edge; the same allocation frequency represents the number of voxels corresponding to the two vertices that are divided into the same function subregion in all individual function subregion division results; performing clustering analysis on the group-level graph structure to obtain a group-level function subregion division result.

3. The method of brain functional subsegmentation of claim 1, wherein, The step of extracting fMRI signals corresponding to each brain region in the target fMRI data to obtain a brain region temporal signal matrix specifically comprises: extracting fMRI signals corresponding to each brain region in the cortical region of the target fMRI data by using a surface-based multi-modal brain region atlas to obtain an fMRI signal time point matrix corresponding to each brain region in the cortical region; extracting fMRI signals corresponding to each brain region in the subcortical region of the target fMRI data by using a Desikan-Killiany atlas template to obtain an fMRI signal time point matrix corresponding to each brain region in the subcortical region. The brain region time signal matrix is determined according to the fMRI signal time point matrix corresponding to each brain region in the cortical region and the fMRI signal time point matrix corresponding to each brain region in the subcortical region.

4. The method of brain functional subsegmentation of claim 1, wherein, The initial spatial feature of the target fMRI data is determined according to the initial time feature and the brain region time signal matrix, and specifically includes: The Pearson correlation coefficient is calculated according to the initial time feature and the brain region time signal matrix; and the Pearson correlation coefficient is taken as the initial spatial feature of the target fMRI data.

5. The method of brain functional subsegmentation of claim 1, wherein, The final similarity matrix is calculated through iterative fusion according to the time similarity matrix and the space similarity matrix, and specifically includes: The time similarity matrix and the space similarity matrix are normalized respectively to obtain a time normalized matrix and a space normalized matrix; The time adjacent voxel similarity matrix and the space adjacent voxel similarity matrix are calculated according to the time similarity matrix and the space similarity matrix respectively; The final similarity matrix is obtained through iterative fusion according to the time adjacent voxel similarity matrix, the space adjacent voxel similarity matrix, the time normalized matrix and the space normalized matrix.

6. The method of brain functional subsegmentation of claim 1, wherein, The individual functional subregion division result of the user is obtained by analyzing the final similarity matrix through the NCut spectral clustering algorithm, and specifically includes: The absolute value conversion is performed on the final similarity matrix to construct an adjacency matrix and a degree matrix; The Laplacian matrix is calculated; The normalized Laplacian matrix corresponding to each voxel is constructed according to the degree matrix and the Laplacian matrix for each voxel; The eigenvalue of the normalized Laplacian matrix corresponding to each voxel is calculated, and all eigenvalues are arranged in ascending order, and the feature vector corresponding to the first k eigenvalues is taken to construct a feature matrix; k is less than the number of voxels; the feature vector matrix is an n x k matrix; n represents the number of voxels; The NCut spectral clustering is performed on the feature vector matrix to obtain the individual functional subregion division result of the user.

7. The method of brain functional subsegmentation of claim 6, wherein, The individual functional subregion division result of the user is obtained, and specifically includes: The individual functional subregion division result of the user is obtained by using the K-means clustering algorithm.

8. A brain functional subregion division apparatus based on the brain functional subregion division method according to any one of claims 1 to 7, characterized by, The brain function subregion division device includes: An original fMRI data acquisition module is configured to acquire original fMRI data of a user. A preprocessing module is configured to preprocess the original fMRI data to obtain target fMRI data. An initial feature extraction module is configured to extract fMRI signals corresponding to each voxel in a region to be divided in the target fMRI data to obtain an initial time feature of the target fMRI data; extract fMRI signals corresponding to each brain region in the target fMRI data to obtain a brain region time signal matrix; and determine an initial spatial feature of the target fMRI data according to the initial time feature and the brain region time signal matrix. A deep feature extraction module is configured to use a trained autoencoder to extract features from the initial time feature and the initial spatial feature of the target fMRI data to obtain deep time features and deep spatial features. The similarity network construction module is configured to construct a similarity network, wherein vertices of the similarity network correspond to voxels in the region to be divided one by one, weights of edges of the similarity network are determined by a time similarity matrix and a space similarity matrix, the time similarity matrix is determined by a Pearson correlation coefficient of depth time features of two voxels connected by an edge, and the space similarity matrix is determined by a Pearson correlation coefficient of depth space features of two voxels connected by an edge. The final similarity matrix calculation module is configured to iteratively fuse the time similarity matrix and the space similarity matrix to calculate a final similarity matrix. The functional subregion division module is configured to, for one user, analyze the final similarity matrix by using an NCut spectral clustering algorithm to obtain a functional subregion division result of the user, and the functional subregion division result includes a plurality of functional subregions.

9. A computer device comprising: 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 brain functional subregion division method in any one of claims 1-7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the brain functional subregion division method in any one of claims 1-7.

Images