Processing method, device and equipment for resting-state brain function magnetic resonance imaging

Abstract

The invention discloses a processing method, device and equipment for resting-state brain function magnetic resonance imaging, and relates to the technical field of computers. According to the method, the time sequence features of each brain region of the user are obtained based on historical resting state brain function magnetic resonance imaging of the user, and then brain map construction is performed according to the time sequence features of each brain region through two different brain map construction modes to obtain brain maps constructed at different angles; when learning is carried out on the basis, the mastery degree of features in resting-state brain function magnetic resonance imaging is improved; then, further diffusing the brain maps constructed at two different angles through a diffusion model, respectively performing feature extraction through two different encoders to obtain brain map features at different angles, and further applying the brain map features to map contrast learning represented by the brain maps, so that the feeling visual field of the brain map contrast learning method is expanded, and the brain map contrast learning efficiency is improved. And the extraction learning ability of the brain map comparison learning method on the intricate interaction between the brain areas is improved.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a method, device and equipment for processing resting-state brain functional magnetic resonance imaging. Background Art

[0002] Currently, brain maps are used to model connections between brain regions and are often used for early diagnosis of brain diseases by providing potential biological information about brain connections.

[0003] In the prior art, Graph Contrastive Learning (GCL) is a self-supervised method that has gradually shown great potential in brain map representation learning and has solved the challenge of difficulty in labeling resting-state functional magnetic resonance imaging.

[0004] However, existing brain map contrastive learning methods cannot fully capture the intricate interactions between brain regions. Summary of the invention

[0005] Based on this, it is necessary to provide a processing method, device and equipment for resting-state brain functional magnetic resonance imaging to address the above-mentioned technical problems.

[0006] The present invention adopts the following technical solutions:

[0007] The present invention provides a method for processing resting-state brain functional magnetic resonance imaging, comprising:

[0008] Obtain the user's historical resting-state functional magnetic resonance imaging, divide the brain regions according to the automatic anatomical labeling atlas, and determine the time series characteristics of each brain region based on the historical resting-state functional magnetic resonance imaging;

[0009] According to the time series characteristics of each brain region, a brain map is constructed using Pearson correlation to obtain the first brain map, and a sparse representation composition method is used to construct the brain map to obtain the second brain map; the first brain map is enhanced by the diffusion model to obtain the third brain map, and the second brain map is enhanced by the diffusion model to obtain the fourth brain map;

[0010] The first brain map and the third brain map are respectively input into the graph isomorphism network, brain map features are extracted from the first brain map to obtain first brain map features, and brain map features are extracted from the third brain map to obtain third brain map features;

[0011] The second brain map and the fourth brain map are respectively input into the Transformer encoder, brain map features are extracted from the second brain map to obtain second brain map features, and brain map features are extracted from the fourth brain map to obtain fourth brain map features;

[0012] A self-contrast loss is determined according to the deviation between the first brain map feature and the third brain map feature and the deviation between the second brain map feature and the fourth brain map feature; a cross-contrast loss is determined according to the deviation between the first brain map feature and the second brain map feature and the deviation between the third brain map feature and the fourth brain map feature; a graph isomorphism network and a Transformer encoder are trained with the goal of minimizing the total loss including the self-contrast loss and the cross-contrast loss, and features of the resting-state functional magnetic resonance imaging to be processed are extracted through the trained graph isomorphism network and the Transformer encoder.

[0013] Optionally, the diffusion model includes a noise unit and a denoising neural network unit;

[0014] The first brain map is enhanced by the diffusion model to obtain the third brain map, and the second brain map is enhanced to obtain the fourth brain map, which specifically includes:

[0015] Obtain the user's historical resting-state functional magnetic resonance imaging and construct a brain map to obtain a sample brain map;

[0016] Multiple rounds of noise addition are performed on the nodes and edges in the sample brain map through the noise unit to obtain a noise map of noisy nodes and noisy edges;

[0017] The noise image is input into a denoising neural network unit, and a sample brain image corresponding to the noise image is predicted by the denoising neural network unit to obtain a predicted sample brain image;

[0018] The denoising neural network unit is trained with the goal of minimizing the deviation between the predicted sample brain map and the sample brain map;

[0019] The first brain map is data enhanced by the trained denoising neural network unit to obtain the third brain map, and the second brain map is data enhanced to obtain the fourth brain map.

[0020] Optionally, each layer of the graph isomorphism network is:

[0021]

[0022] Among them, φ (l+1) Represents the multi-layer perceptron function, ∈ (l+1) is a trainable parameter, represents the feature vector of node v in the lth layer, N(v) represents the set of neighbor nodes of node v, Represents the feature vector of the l-th layer node u, where the node is a node in the brain map and each node corresponds to a brain region.

[0023] Optionally, the number of layers of the graph isomorphism network is 2;

[0024] The extracting the first brain map feature from the first brain map to obtain the first brain map feature specifically includes:

[0025] The updated node features are obtained by extracting brain region features from each node in the first brain map using the following formula:

[0026]

[0027] A first mind map feature is obtained by averaging the updated node features of each node in the first mind map;

[0028] in, is the updated node feature, ρ is the ReLU activation function, W (0) represents the weight matrix of the first layer of the graph isomorphism network in the multilayer perceptron, W (1) It represents the weight matrix of the second layer of the graph isomorphism network in the multilayer perceptron, I is the identity matrix, A is the edge set of the first brain map, and X is the node set of the first brain map.

[0029] Optionally, determining the total loss specifically includes:

[0030] The self-contrastive loss is determined according to the deviation between the first brain map feature and the third brain map feature and the deviation between the second brain map feature and the fourth brain map feature by the following formula:

[0031]

[0032] The cross-contrast loss is determined according to the deviation between the first brain map feature and the second brain map feature and the deviation between the third brain map feature and the fourth brain map feature by the following formula:

[0033]

[0034] The total loss is determined by:

[0035]

[0036] Among them, P a1 is the first brain map feature, P a2 is the third brain map feature, P b1 is the second brain map feature, P b2 is the fourth brain map feature, Φ(P a1 ,P a2 ) is the deviation between the first brain map feature and the third brain map feature, Φ(P b1 ,P b2 ) is the deviation between the second brain map feature and the fourth brain map feature, is the self-contrast loss, Φ(P a1 ,P b1 ) is the deviation between the first brain map feature and the second brain map feature, Φ(Pa2 ,P b2 ) is the deviation between the third brain map feature and the fourth brain map feature, is the cross-contrast loss, λ1 and λ2 are adjustable hyperparameters, is the total loss.

[0037] Optionally, the deviation loss between the predicted sample brain map and the sample brain map is determined by the following formula:

[0038]

[0039] The total loss is determined by:

[0040]

[0041] in, is the deviation loss between the predicted sample brain map and the sample brain map, x i is the node feature of the sample brain map, To predict the node features of the sample brain map, a i,j is the edge feature of the sample brain map, is to predict the edge features of the sample brain map, N is the total number of nodes in the sample brain map, is the self-contrastive loss, is the cross-contrastive loss, and λ1, λ2, and λ3 are adjustable hyperparameters.

[0042] The present invention provides a processing device for resting-state brain functional magnetic resonance imaging, comprising:

[0043] An acquisition module, used for acquiring a user's resting-state functional magnetic resonance imaging, and determining the time series characteristics of each brain region based on the resting-state functional magnetic resonance imaging according to the brain region division of the automatic anatomical labeling atlas;

[0044] A data enhancement module is used to construct a brain map using Pearson correlation according to the time series characteristics of each brain region to obtain a first brain map, and to construct a brain map using a sparse representation composition method to obtain a second brain map; to perform data enhancement on the first brain map using a diffusion model to obtain a third brain map, and to perform data enhancement on the second brain map using a diffusion model to obtain a fourth brain map;

[0045] A first feature extraction module is used to input the first brain map and the third brain map into a graph isomorphism network including a double-layer feature extraction layer, perform brain map feature extraction on the first brain map to obtain a first brain map feature, and perform brain map feature extraction on the third brain map to obtain a third brain map feature;

[0046] A second feature extraction module is used to input the second brain map and the fourth brain map into the Transformer encoder respectively, perform brain map feature extraction on the second brain map to obtain second brain map features, and perform brain map feature extraction on the fourth brain map to obtain fourth brain map features;

[0047] The training application module is used to determine the self-contrast loss according to the deviation between the first brain map feature and the third brain map feature and the deviation between the second brain map feature and the fourth brain map feature; determine the cross-contrast loss according to the deviation between the first brain map feature and the second brain map feature and the deviation between the third brain map feature and the fourth brain map feature; train the graph isomorphism network and the Transformer encoder with the goal of minimizing the total loss including the self-contrast loss and the cross-contrast loss, and extract features from the resting-state brain functional magnetic resonance imaging to be processed through the trained graph isomorphism network and the Transformer encoder.

[0048] The present invention provides a computer-readable storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the processing method of resting-state brain functional magnetic resonance imaging is implemented.

[0049] The present invention 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 implements the above-mentioned resting state brain functional magnetic resonance imaging processing method when executing the program.

[0050] At least one of the above technical solutions adopted by the present invention can achieve the following beneficial effects:

[0051] The present invention first obtains the time series characteristics of each brain region of the user based on the user's historical resting-state functional magnetic resonance imaging, and then constructs a brain map according to the time series characteristics of each brain region through two different brain map construction methods, namely Pearson correlation and sparse representation composition, to obtain brain maps constructed at different angles. When learning based on this, the mastery of the features in the resting-state functional magnetic resonance imaging is improved; then, the brain maps constructed at two different angles are further diffused through a diffusion model, and feature extraction is performed respectively through two different encoders, namely a graph isomorphism network and a Transformer encoder, to obtain brain map features at different angles, which are further applied to graph contrast learning represented by brain maps, thereby expanding the perception field of the brain map contrast learning method and improving the ability of the brain map contrast learning method to extract and learn the complex interactions between brain regions. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0053] Figure 1 A schematic flow chart of a method for processing resting-state brain functional magnetic resonance imaging provided by the present invention;

[0054] Figure 2 A schematic diagram of a brain map construction, data enhancement and model training framework provided by the present invention;

[0055] Figure 3 A schematic diagram of the training and application process of a resting-state functional magnetic resonance imaging processing method provided by the present invention;

[0056] Figure 4 A schematic diagram of a processing device for resting-state brain functional magnetic resonance imaging provided by the present invention;

[0057] Figure 5 A schematic diagram of a computer device for implementing a processing method for resting-state brain functional magnetic resonance imaging provided by the present invention. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0059] In order to fully capture the intricate interactions between brain regions, this paper proposes a new heterogeneous brain map contrast learning framework HeBrainGCL, which uses heterogeneous brain map pairs and heterogeneous encoders to capture more comprehensive and complementary biological information to assist doctors in brain disease detection. Specifically, HeBrainGCL uses various types of brain maps generated by the Graph Diffusion Model and their corresponding enhancements to establish two heterogeneous contrast pair views. Based on this, two heterogeneous encoders are used to form a dual contrast mode: self-contrast and cross-contrast. Among them, self-contrast independently performs classical contrast learning on each encoder, while cross-contrast promotes knowledge interaction between different models. In this way, HeBrainGCL extracts more and more reasonable brain connection information between brain regions.

[0060] The technical solutions provided by various embodiments of the present invention are described in detail below in conjunction with the accompanying drawings.

[0061] Figure 1 The present invention is a flowchart of a method for processing resting state brain functional magnetic resonance imaging, which specifically includes the following steps:

[0062] S101: Obtain the user's resting-state functional magnetic resonance imaging, divide the brain regions according to the automatic anatomical labeling atlas, and determine the time series characteristics of each brain region based on the resting-state functional magnetic resonance imaging.

[0063] S102: constructing a brain map using Pearson correlation according to the time series characteristics of each brain region to obtain a first brain map, and constructing a brain map using a sparse representation composition method to obtain a second brain map; performing data enhancement on the first brain map using a diffusion model to obtain a third brain map, and performing data enhancement on the second brain map using a diffusion model to obtain a fourth brain map.

[0064] S103: inputting the first brain map and the third brain map into the graph isomorphism network respectively, performing brain map feature extraction on the first brain map to obtain a first brain map feature, and performing brain map feature extraction on the third brain map to obtain a third brain map feature.

[0065] S104: inputting the second brain map and the fourth brain map into the Transformer encoder respectively, performing brain map feature extraction on the second brain map to obtain second brain map features, and performing brain map feature extraction on the fourth brain map to obtain fourth brain map features.

[0066] S105: Determine a self-contrast loss based on the deviation between the first brain map feature and the third brain map feature and the deviation between the second brain map feature and the fourth brain map feature; determine a cross-contrast loss based on the deviation between the first brain map feature and the second brain map feature and the deviation between the third brain map feature and the fourth brain map feature; train the graph isomorphism network and the Transformer encoder with the goal of minimizing the total loss including the self-contrast loss and the cross-contrast loss, and extract features from the resting-state functional magnetic resonance imaging to be processed through the trained graph isomorphism network and the Transformer encoder.

[0067] For the convenience of explanation, the following description is only based on the server as the execution subject. The server mentioned in the present invention can be a server set up on the business platform, or a device such as a desktop computer, a notebook computer, etc. that can execute the solution of the present invention.

[0068] Generally, during the examination of brain diseases, the user's resting-state functional magnetic resonance imaging can usually be obtained first, and then processed by the server of the business platform, so as to improve the degree of representation of brain diseases by the processed data, so as to further assist doctors in examinations.

[0069] Based on this, in one or more embodiments of the present invention, the server may first obtain the user's resting state brain functional magnetic resonance imaging and pre-process it.

[0070] Specifically, the server can first perform slice time correction (removing the first p time points of the historical resting-state functional MRI) and head motion correction on the user's historical resting-state functional MRI, and remove the influence of ventricle, white matter signals and head motion high-order effects in the corrected historical resting-state functional MRI to obtain the historical resting-state functional MRI after preliminary preprocessing.

[0071] Then, the historical resting-state functional magnetic resonance imaging after preliminary preprocessing was aligned to the standard space, and the historical resting-state functional magnetic resonance imaging after alignment was subjected to a 0.01-0.1 Hz time bandpass filter to reduce the influence of heartbeat and breathing, so as to obtain the final preprocessed historical resting-state functional magnetic resonance imaging.

[0072] Finally, the brain was divided into multiple brain regions according to the automatic anatomical labeling atlas, and the average time series characteristics of each brain region were extracted based on the historical resting-state functional magnetic resonance imaging after the final preprocessing as the time series characteristics of each brain region.

[0073] For example, the server can use the public Autism Brain Imaging Data Exchange (ABIDE) dataset and select the data from the two largest sites, namely New York University (NYU) and the University of Michigan (UM); the NYU dataset has a total of 184 subjects, including 79 autistic patients and 105 healthy controls; the UM dataset has a total of 145 subjects, including 68 autistic patients and 77 healthy controls. Read and convert the format of the resting-state brain functional magnetic resonance imaging data collected by the magnetic resonance device. Except for the first 10 unstable time points of each subject; perform head movement and time layer correction; regress out the influence of ventricles, white matter, and global signals; align to the standard space through the Montreal Neurological Institute template; filter with a 0.01-0.1HZ bandpass filter to reduce the influence of heartbeat and breathing.

[0074] Furthermore, the server may select a standardized brain partition template. For example, the AAL template is selected to divide the brain into 116 brain regions, including 90 cerebral brain regions and 26 cerebellar brain regions. The 116 brain regions represent 116 nodes of the functional brain network. The average time series of each brain region is extracted: S = [s1, s2, …, s i ,…,s n ]∈R t×n , where s i ∈R t represents the time series characteristics of the ith brain region, t=175 represents the length of the time series, and n=116 represents the number of brain regions.

[0075] After obtaining the time series characteristics of each brain region, the server may use different composition methods to process the time series characteristics of each brain region. In one or more embodiments of the present invention, PC (Pearson Correlation Coefficient) composition and SR (Sparse Representation for Reconstruction) composition may be used to construct a brain map.

[0076] For PC mapping, the server may use Pearson correlation to construct the brain map, thereby obtaining a first brain map. For SR mapping, the server may use sparse representation mapping method to construct the brain map, thereby obtaining a second brain map.

[0077] Based on this, in order to solve the problem of data scarcity, the server can further perform data enhancement on the first brain map through a diffusion model to obtain a third brain map, and perform data enhancement on the second brain map to obtain a fourth brain map.

[0078] Specifically, in order to avoid losing valuable information in the original signal, the present invention chooses to enhance at the graph level rather than at the blood oxygen level dependent (BOLD) signal level. This involves using a graph diffusion model to simultaneously adjust the topological structure and node features of the brain map to improve its expressiveness. The diffusion enhancement module consists of two parts: a noise unit and a denoising neural network unit.

[0079] When training the diffusion model, the user's historical resting-state functional magnetic resonance imaging can be obtained first and a brain map can be constructed to obtain a sample brain map. Then, multiple rounds of noise are added to the nodes and edges in the sample brain map through the noise unit to obtain a noise map of noisy nodes and noisy edges. Then, the noise map is input into the denoising neural network unit, and the sample brain map corresponding to the noise map is predicted by the denoising neural network unit to obtain a predicted sample brain map. Finally, the denoising neural network unit is trained with the goal of minimizing the deviation between the predicted sample brain map and the sample brain map.

[0080] For the noise unit: the noise unit can be represented by the Markov transfer matrix Q t Definition (which can be understood as a matrix that can be continuously accumulated). In one or more embodiments of the present invention, the sample brain map G can be diffused on each node and edge feature separately. For any node (or edge) in the diffusion step T, the transition probability can be represented by the matrix and definition.

[0081] Where t∈{1,.,T} represents the number of diffusion steps, m corresponds to the states of nodes and edges in step t-1, and n corresponds to the states of nodes and edges in diffusion step t. is the transition probability of the node in the diffusion step t, is the node after diffusion step t, is the transition probability of edge diffusion step t, a t is the edge after diffusion step t.

[0082] After adding noise, G t =(X t ,A t ), each new noisy node and edge can be obtained by the following formula:

[0083] in, G t represents the final image after adding noise, X t is the node feature, A t The edge feature.

[0084] When considering an undirected graph, we can apply the noise only to the upper triangular part of the matrix and then be symmetric with the matrix.

[0085] According to the above description, Q t The process of adding noise to the data can be controlled, Q t The choice of is not arbitrary. It needs to satisfy two conditions: first, Q t The rows of must sum to 1 to preserve the probability mass. Second, as t gets larger, Q t It should be able to converge to a known stationary distribution. A common practice is to use a doubly random matrix (rows and columns sum to 1) to define the transfer matrix Q t :

[0086] Among them, I is the identity matrix, K represents the last dimension of node / edge features, 1 K is a column vector with all elements equal to 1, α t Following the cosine schedule: β t =1-α t α t Here s is a very small value. It can be determined by the following formula:

[0087] For the denoising neural network unit: The second component of the GDA module is a denoising neural network parameterized by θ It is based on the noise graph G t =(X t ,At ) as input, and through training it is able to predict a clean sample brain map G that is close to the real map.

[0088] For training In one or more embodiments of the present invention, the sample brain map can be predicted based on The diffusion loss is optimized by the mean square error between the real sample brain map G with nodes and edges. The diffusion loss can be defined by the following formula:

[0089] Among them, x i It represents the node features of the real brain map. It represents the node features of the predicted sample brain map, a i,j It represents the edge features of the real sample brain map. It represents the edge features of the predicted sample brain map.

[0090] Once a denoising neural network unit is trained, it can be used to sample new images. To do this, the network predictions are used. Estimate the back diffusion iteration P θ (G t-1 |G t ). This distribution can be modeled as the product of nodes and edges:

[0091] In the formula, Indicates reverse diffusion of nodes. Indicates reverse diffusion of the edges.

[0092] Each term can be calculated based on the network prediction:

[0093] This formula can be used to predict the node graph in the previous step after adding noise. It should be noted here that:

[0094] Similarly, for edge features

[0095] This formula can be used to predict the edge graph in the previous step after adding noise. These distributions are used to t-1 Sampling, G t-1 It will be used as the input of the denoising network and iterated t times until the generated predicted sample brain map G' is obtained.

[0096] Figure 2Figure 1 is a schematic diagram of a brain map construction, data enhancement and model training framework in the present invention. Figure 2 It can be seen that after PC composition and SR composition are performed on the user's historical resting-state functional magnetic resonance imaging, data enhancement can be further performed, and a brain map feature extraction comparative learning process can be performed through different models based on the original data before enhancement and the original data after enhancement, thereby improving the model's brain map feature extraction ability.

[0097] After enhancing the PC composition data, the server can further achieve high-quality feature extraction through the GIN network. Compared with the traditional graph neural network (GNN), the GIN network uses a different aggregator, which uses a summation operation to aggregate neighborhood node features. In addition, the GIN network introduces a multilayer perceptron (MLP) to perform nonlinear transformations on the aggregated neighborhood and original node features. Each layer of GIN is:

[0098] In the formula, φ (l+1) Represents the multi-layer perceptron function, ∈ (l+1) is a trainable parameter, represents the feature vector of node v in the lth layer, N(v) represents the set of neighbor nodes of node v, Represents the feature vector of the l-th layer node u, where the node is a node in the brain map and each node corresponds to a brain region.

[0099] In one or more embodiments of the present invention, the number of GIN layers can be set to 2. Taking a single branch as an example, the brain map can be input into GIN for feature learning. Therefore, after two stacked GIN layers, taking the first brain map feature extraction to obtain the first brain map feature as an example, the brain region feature extraction of each node in the first brain map can be performed through the following formula to obtain the updated node feature:

[0100] In the formula, is the updated node feature, ρ is the ReLU activation function, W (0) represents the weight matrix of the first layer of the graph isomorphism network in the multilayer perceptron, W (1) It represents the weight matrix of the second layer of the graph isomorphism network in the multi-layer perceptron, I is the identity matrix, A represents the edge features, and X represents the node features.

[0101] Then, the updated node features of each node in the first brain map can be averaged to obtain the first brain map feature P a1 For the third brain map, the third brain map feature P can be obtained in the same way. a2 , I will not go into details here.

[0102] After the SR composition data is enhanced, in one or more embodiments of the present invention, a Transformer encoder can be used to extract the features of the data. Compared with the traditional Transformer encoder, the position encoding information can be removed because it is redundant to use position encoding for the matrix data of the present invention. The Transformer encoder mainly includes several main components: position feedforward network, scaled dot product attention mechanism, multi-head attention mechanism. The input data is processed through each encoder layer, and the attention weight of each layer is collected, and then flattened and further processed through the linear layer. Finally, the corresponding second brain map feature P is obtained. b1 and the fourth brain map feature P b2 .

[0103] Contrastive loss function: Given the last four brain map features P a1 ,P a2 ,P b1 ,P b2 , where P a1 ,P a2 Different views from the same subject, P b1 ,P b2 Different views from the same subject are assumed to show similarity in the newly mapped embedding space. In one or more embodiments of the present invention, the loss function is divided into a self-contrast loss function and a hypercontrast loss function. The self-contrast loss function compares the original data and the enhanced data from the same composition method. For example, for PC composition, the goal is to minimize the two output vectors P a1 and P a2 The negative cosine similarity between them can be determined by the following formula according to the deviation between the first brain map feature and the third brain map feature and the deviation between the second brain map feature and the fourth brain map feature:

[0104] In the formula, Φ(P a1 ,P a2 ) is the deviation between the first brain map feature and the third brain map feature, Φ(P b1 ,P b2 ) is the deviation between the second brain map feature and the fourth brain map feature, is the self-contrast loss.

[0105] For SR mapping, the goal is to minimize the two output vectors P b1 and P b2 The negative cosine similarity between , can be determined by the following formula according to the deviation between the first brain map feature and the second brain map feature and the deviation between the third brain map feature and the fourth brain map feature:

[0106] In the formula, Φ(P a1 ,P b1 ) is the deviation between the first brain map feature and the second brain map feature, Φ(P a2 ,P b2 ) is the deviation between the third brain map feature and the fourth brain map feature, is the cross-contrast loss.

[0107] Based on the above, if the diffusion process and contrastive learning process are trained separately, the total loss can be determined by the following formula:

[0108] If the overall training is performed, the total loss can be determined by the following formula:

[0109] Where λ1, λ2 and λ3 are adjustable hyperparameters. is the total loss.

[0110] After completing the training of all models, the server can extract features from the resting-state functional magnetic resonance imaging to be processed through the trained graph isomorphism network and Transformer encoder. Figure 3 The present invention is a schematic diagram of a training and application process of a processing method for resting-state functional magnetic resonance imaging of the brain.

[0111] based on Figure 1 The processing method of resting-state functional magnetic resonance imaging shown in the present invention first obtains the time series characteristics of each brain region of the user based on the user's historical resting-state functional magnetic resonance imaging, and then constructs a brain map according to the time series characteristics of each brain region through two different brain map construction methods, namely Pearson correlation and sparse representation composition, to obtain brain maps constructed at different angles. When learning based on this, the mastery of the features in resting-state functional magnetic resonance imaging is improved; then, the brain maps constructed at two different angles are further diffused through a diffusion model, and feature extraction is performed respectively through two different encoders, namely a graph isomorphism network and a Transformer encoder, to obtain brain map features at different angles, which are further applied in graph contrast learning represented by brain maps, thereby expanding the perception field of the brain map contrast learning method and improving the ability of the brain map contrast learning method to extract and learn the complex interactions between brain regions.

[0112] When applying the processing method of resting state brain functional magnetic resonance imaging provided by the present invention, it is not necessary to Figure 1 The steps are executed in the order shown. The specific execution order of the steps can be determined according to needs, and the present invention does not limit this.

[0113] The above is a method for processing resting state brain functional magnetic resonance imaging provided by one or more embodiments of the present invention. Based on the same idea, the present invention also provides a corresponding device for processing resting state brain functional magnetic resonance imaging, such as Figure 4 shown.

[0114] Figure 4 A schematic diagram of a processing device for resting-state brain functional magnetic resonance imaging provided by the present invention, comprising:

[0115] An acquisition module 201 is used to acquire the user's historical resting-state functional magnetic resonance imaging, and determine the time series characteristics of each brain region based on the historical resting-state functional magnetic resonance imaging according to the brain region division of the automatic anatomical labeling atlas;

[0116] The data enhancement module 202 is used to construct a brain map using Pearson correlation according to the time series characteristics of each brain region to obtain a first brain map, and to construct a brain map using a sparse representation composition method to obtain a second brain map; to perform data enhancement on the first brain map using a diffusion model to obtain a third brain map, and to perform data enhancement on the second brain map using a diffusion model to obtain a fourth brain map;

[0117] A first feature extraction module 203 is used to input the first brain map and the third brain map into a graph isomorphism network including a double-layer feature extraction layer, perform brain map feature extraction on the first brain map to obtain a first brain map feature, and perform brain map feature extraction on the third brain map to obtain a third brain map feature;

[0118] A second feature extraction module 204 is used to input the second brain map and the fourth brain map into the Transformer encoder respectively, perform brain map feature extraction on the second brain map to obtain a second brain map feature, and perform brain map feature extraction on the fourth brain map to obtain a fourth brain map feature;

[0119] The training application module 205 is used to determine the self-contrast loss according to the deviation between the first brain map feature and the third brain map feature and the deviation between the second brain map feature and the fourth brain map feature; determine the cross-contrast loss according to the deviation between the first brain map feature and the second brain map feature and the deviation between the third brain map feature and the fourth brain map feature; train the graph isomorphism network and the Transformer encoder with the goal of minimizing the total loss including the self-contrast loss and the cross-contrast loss, and extract features from the resting-state brain functional magnetic resonance imaging to be processed through the trained graph isomorphism network and the Transformer encoder.

[0120] The specific definition of the processing device of resting state brain functional magnetic resonance imaging can be found in the definition of the processing method of resting state brain functional magnetic resonance imaging above, which will not be repeated here. Each module in the above-mentioned processing device of resting state brain functional magnetic resonance imaging can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0121] The present invention also provides a computer-readable storage medium, which stores a computer program, which can be used to execute the above Figure 1 Provided are methods for processing resting-state functional magnetic resonance imaging of the brain.

[0122] The present invention also provides Figure 5 The structural diagram of the computer device shown in FIG. Figure 5 As shown in the figure, at the hardware level, the computer device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory, and may also include other hardware required for the business. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to achieve the above Figure 1 Provided are methods for processing resting-state functional magnetic resonance imaging of the brain.

[0123] 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, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention 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 or optical memory, etc. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0124] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, 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 the present invention.

Claims

1. A method for processing resting state brain functional magnetic resonance imaging, characterized in that: include: Obtain the user's historical resting-state functional magnetic resonance imaging, divide the brain regions according to the automatic anatomical labeling atlas, and determine the time series characteristics of each brain region based on the historical resting-state functional magnetic resonance imaging; According to the time series characteristics of each brain region, a brain map is constructed using Pearson correlation to obtain the first brain map, and a sparse representation composition method is used to construct the brain map to obtain the second brain map; the first brain map is enhanced by the diffusion model to obtain the third brain map, and the second brain map is enhanced by the diffusion model to obtain the fourth brain map; The first brain map and the third brain map are respectively input into the graph isomorphism network, brain map features are extracted from the first brain map to obtain first brain map features, and brain map features are extracted from the third brain map to obtain third brain map features; The second brain map and the fourth brain map are respectively input into the Transformer encoder, brain map features are extracted from the second brain map to obtain second brain map features, and brain map features are extracted from the fourth brain map to obtain fourth brain map features; A self-contrast loss is determined according to the deviation between the first brain map feature and the third brain map feature and the deviation between the second brain map feature and the fourth brain map feature; a cross-contrast loss is determined according to the deviation between the first brain map feature and the second brain map feature and the deviation between the third brain map feature and the fourth brain map feature; a graph isomorphism network and a Transformer encoder are trained with the goal of minimizing the total loss including the self-contrast loss and the cross-contrast loss, and features of the resting-state functional magnetic resonance imaging to be processed are extracted through the trained graph isomorphism network and the Transformer encoder.

2. The method for processing resting state brain functional magnetic resonance imaging according to claim 1, characterized in that: The diffusion model includes a noise unit and a denoising neural network unit; The first brain map is enhanced by the diffusion model to obtain the third brain map, and the second brain map is enhanced to obtain the fourth brain map, which specifically includes: Obtain the user's historical resting-state functional magnetic resonance imaging and construct a brain map to obtain a sample brain map; Multiple rounds of noise addition are performed on the nodes and edges in the sample brain map through the noise unit to obtain a noise map of noisy nodes and noisy edges; The noise image is input into a denoising neural network unit, and a sample brain image corresponding to the noise image is predicted by the denoising neural network unit to obtain a predicted sample brain image; The denoising neural network unit is trained with the goal of minimizing the deviation between the predicted sample brain map and the sample brain map; The first brain map is data enhanced by the trained denoising neural network unit to obtain the third brain map, and the second brain map is data enhanced to obtain the fourth brain map.

3. The method for processing resting state brain functional magnetic resonance imaging according to claim 1, characterized in that: Each layer of the graph isomorphism network is: Among them, φ (l+1) Represents the multi-layer perceptron function, ∈ (l+1) is a trainable parameter, represents the feature vector of node v in the lth layer, N(v) represents the set of neighbor nodes of node v, Represents the feature vector of the l-th layer node u, where the node is a node in the brain map and each node corresponds to a brain region.

4. The method for processing resting state brain functional magnetic resonance imaging according to claim 3, characterized in that: The number of layers of the graph isomorphism network is 2; The extracting the first brain map feature from the first brain map to obtain the first brain map feature specifically includes: The updated node features are obtained by extracting brain region features from each node in the first brain map using the following formula: A first mind map feature is obtained by averaging the updated node features of each node in the first mind map; in, is the updated node feature, ρ is the ReLU activation function, W (0) represents the weight matrix of the first layer of the graph isomorphism network in the multi-layer perceptron, W (1) represents the weight matrix of the second layer of the graph isomorphism network in the multilayer perceptron, I is the identity matrix, A represents the edge of the first brain map, and X represents the node of the first brain map.

5. The method for processing resting state brain functional magnetic resonance imaging according to claim 1, characterized in that: Determine the total loss, including: The self-contrastive loss is determined according to the deviation between the first brain map feature and the third brain map feature and the deviation between the second brain map feature and the fourth brain map feature by the following formula: The cross-contrast loss is determined according to the deviation between the first brain map feature and the second brain map feature and the deviation between the third brain map feature and the fourth brain map feature by the following formula: The total loss is determined by: Among them, P a1 is the first brain map feature, P a2 is the third brain map feature, P b1 is the second brain map feature, P b2 is the fourth brain map feature, Φ(P a1 , P a2 ) is the deviation between the first brain map feature and the third brain map feature, Φ(P b1 , P b2 ) is the deviation between the second brain map feature and the fourth brain map feature, is the self-contrast loss, Φ(P a1 , P b1 ) is the deviation between the first brain map feature and the second brain map feature, Φ(P a2 , P b2 ) is the deviation between the third brain map feature and the fourth brain map feature, is the cross-contrast loss, λ1 and λ2 are adjustable hyperparameters, is the total loss.

6. The method for processing resting state brain functional magnetic resonance imaging according to claim 2, characterized in that: The deviation loss between the predicted sample brain map and the sample brain map is determined by the following formula: The total loss is determined by: in, is the deviation loss between the predicted sample brain map and the sample brain map, x i is the node feature of the sample brain map, To predict the node features of the sample brain map, a i，j is the edge feature of the sample brain map, is to predict the edge features of the sample brain map, N is the total number of nodes in the sample brain map, is the self-contrastive loss, is the cross-contrastive loss, and λ1, λ2, and λ3 are adjustable hyperparameters.

7. A processing device for resting state brain functional magnetic resonance imaging, characterized in that: include: An acquisition module, used to acquire the user's historical resting-state functional magnetic resonance imaging, and determine the time series characteristics of each brain region based on the historical resting-state functional magnetic resonance imaging according to the brain region division of the automatic anatomical labeling atlas; A data enhancement module is used to construct a brain map using Pearson correlation according to the time series characteristics of each brain region to obtain a first brain map, and to construct a brain map using a sparse representation composition method to obtain a second brain map; to perform data enhancement on the first brain map using a diffusion model to obtain a third brain map, and to perform data enhancement on the second brain map using a diffusion model to obtain a fourth brain map; A first feature extraction module is used to input the first brain map and the third brain map into a graph isomorphism network including a double-layer feature extraction layer, perform brain map feature extraction on the first brain map to obtain a first brain map feature, and perform brain map feature extraction on the third brain map to obtain a third brain map feature; A second feature extraction module is used to input the second brain map and the fourth brain map into the Transformer encoder respectively, perform brain map feature extraction on the second brain map to obtain second brain map features, and perform brain map feature extraction on the fourth brain map to obtain fourth brain map features; The training application module is used to determine the self-contrast loss according to the deviation between the first brain map feature and the third brain map feature and the deviation between the second brain map feature and the fourth brain map feature; determine the cross-contrast loss according to the deviation between the first brain map feature and the second brain map feature and the deviation between the third brain map feature and the fourth brain map feature; train the graph isomorphism network and the Transformer encoder with the goal of minimizing the total loss including the self-contrast loss and the cross-contrast loss, and extract features from the resting-state brain functional magnetic resonance imaging to be processed through the trained graph isomorphism network and the Transformer encoder.

8. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

9. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method according to any one of claims 1 to 6 is implemented.

Images