Intelligent identification method and device for rs-fMRI data

Abstract

The invention provides an rs-fMRI data intelligent identification method and device, and the method comprises the following steps: S1, carrying out the preprocessing of rs-fMRI data, and enabling the rs-fMRI data to comprise a plurality of time sequences; and S2, inputting the preprocessed rs-fMRI data into a pre-trained rs-fMRI identification model, and obtaining a classification label of the rs-fMRI data. The deep learning technology is adopted, especially the time-space double-path cross attention network based on deep learning is adopted, the analysis efficiency is high, the adaptability is high, and the rs-fMRI data can be accurately recognized.

Description

Technical Field

[0001] The present invention relates to the field of medical image data processing technology, and in particular to a method and device for intelligent recognition of rs-fMRI data based on a spatiotemporal dual-path cross-attention mechanism. Background Art

[0002] Functional magnetic resonance imaging (fMRI), a non-invasive medical imaging technique, has become an important tool for studying human brain activity. Resting-state functional MRI (rs-fMRI) can capture spontaneous brain activity while subjects are at rest and reveal natural connectivity patterns between brain regions, which constitute functional connectivity networks. Although rs-fMRI provides valuable data for the study of major depressive disorder (MDD), effectively extracting features from rs-fMRI data for analysis to accurately classify samples into predefined categories and achieve intelligent recognition remains a challenge.

[0003] Traditional methods usually rely on manual feature extraction and statistical analysis, and have the following limitations: 1. Limited feature extraction capabilities: Manual feature extraction often relies on existing prior knowledge and cannot fully capture the complex patterns and dynamic changes in rs-fMRI data; 2. Low analysis efficiency: Traditional methods are inefficient when processing large-scale data sets; 3. Poor adaptability to individual differences: Traditional methods often ignore the physiological and behavioral differences between different individuals, resulting in limited model generalization capabilities; 4. Insufficient recognition accuracy: Traditional methods still have much room for improvement in the accuracy and reliability of intelligent recognition of rs-fMRI data.

[0004] Despite recent significant progress in fields such as image processing and speech recognition, deep learning applications in medical image analysis, particularly in rs-fMRI data processing, remain exploratory. Deep learning models, with their powerful feature extraction and pattern recognition capabilities, have the potential to overcome the shortcomings of traditional methods. However, designing deep learning models that can adapt to the characteristics of rs-fMRI data remains an important research direction and challenge.

[0005] Therefore, it is urgent to develop an intelligent recognition method and device for rs-fMRI data based on deep learning. Summary of the Invention

[0006] (1) Technical issues to be solved

[0007] In view of the problems existing in the above-mentioned technologies, the present invention aims to address them, at least to a certain extent. To this end, the present invention aims to propose a method and device for intelligent recognition of rs-fMRI data based on deep learning, which has high analysis efficiency, strong adaptability, and can accurately recognize rs-fMRI data.

[0008] (2) Technical solution

[0009] In order to achieve the above objectives, the main technical solutions adopted by the present invention include:

[0010] In a first aspect, the present invention provides a method for intelligent recognition of rs-fMRI data, comprising the following steps:

[0011] S1. Preprocessing rs-fMRI data. rs-fMRI data contains multiple time series.

[0012] S2, input the pre-processed rs-fMRI data into the pre-trained rs-fMRI recognition model to obtain the classification label of the rs-fMRI data;

[0013] The rs-fMRI recognition model includes a temporal feature extraction path, a spatial feature extraction path, and a cross-attention module. In the temporal feature extraction path, multiple time series are input into the long short-term memory (LSTM) network, and the hidden state sequence corresponding to each time series is output. The perturbation position code corresponding to each time series is generated according to the position index in the time series and the random perturbation noise. The enhanced features corresponding to each time series obtained by adding the hidden state sequence and the perturbation position code corresponding to the same time series are output as the embedded features of the temporal path. In the spatial feature extraction path, the Gramian angle field is used to convert each time series into a GAF image, and the structural similarity index between GAF images is calculated. The whole-brain functional connectivity map is constructed based on the structural similarity index between GAF images. The whole-brain functional connectivity map is input into the graph convolutional network, and the embedded features of the spatial path are output. The embedded features of the temporal path and the embedded features of the spatial path are input into the cross-attention module, and the classification label of the rs-fMRI data is output.

[0014] Optionally, preprocess the rs-fMRI data, including:

[0015] S01, perform spatial normalization on rs-fMRI data;

[0016] S02. truncating multiple time series according to a preset sequence length;

[0017] S03. Fill in missing values in multiple time series.

[0018] Optionally, generating a disturbance position code corresponding to each time series according to the position index in the time series and the random disturbance noise includes:

[0019]

[0020]

[0021] Where PPE is the perturbation position code corresponding to the time series; pos is the position index in the time series; p is the random disturbance noise; d e is the dimension of the embedded feature; i is the count of the time point; i%2 is i divisible by 2; dim is the dimension; is the size of PPE, and m is the time point.

[0022] Optionally, each time series is converted into a GAF image using the Gramian angle field, which is expressed as follows:

[0023]

[0024] Where, X c is the time series of the cth region of interest of rs-fMRI; χ c For X c GAF image; I is the identity matrix; T is the transpose operation of the matrix.

[0025] Optionally, a whole-brain functional connectivity map is constructed based on the structural similarity index between GAF images, and the formula is expressed as:

[0026]

[0027] Where, is the structural similarity index between GAF images; and They are the GAF images of the c1 channel and the GAF images of the c2 channel respectively; and They are and The average value of and They are and The standard deviation of for and covariance of ; δ1 and δ2 are both constants.

[0028] Optionally, the spatial feature extraction path further includes: performing a sparse processing on the whole-brain functional connectivity map according to a preset threshold to obtain an adjacency matrix, constructing a sparse whole-brain functional connectivity map according to the adjacency matrix, inputting the sparse whole-brain functional connectivity map into a graph convolutional network, and outputting the embedded features of the spatial path;

[0029] The formula of the adjacency matrix is expressed as:

[0030]

[0031] Where, is the adjacency matrix; is the structural similarity index between GAF images; τ is a pre-set threshold.

[0032] Optionally, the crisscross attention module includes a dual-path crisscross attention layer and a two-dimensional convolutional layer;

[0033] The embedded features of the temporal path and the embedded features of the spatial path are input into the dual-path cross attention layer, and the attention results of the temporal path and the attention results of the spatial path are output. The attention results of the temporal path and the attention results of the spatial path are integrated into a tensor through the outer product, and the tensor is input into the two-dimensional convolutional layer to output the classification label of the rs-fMRI data.

[0034] Optionally, the attention results of the temporal path and the attention results of the spatial path are integrated into a tensor through the outer product, which is expressed as follows:

[0035] P = ρ S ×ρ T

[0036] Where P is a tensor, R is the size of PPE, m is the time point, d e is the dimension of the embedded feature; ρ S is the attention result of the time path; ρ T is the attention result of the spatial path;

[0037] The tensor is input into the two-dimensional convolutional layer, and the classification label of the rs-fMRI data is output. The formula is expressed as:

[0038] Z=Conv 2D (P)

[0039] Where Z is the classification label of rs-fMRI data, R is the size of PPE, m is the time point, d e is the dimension of the embedded feature; Conv 2D It is a two-dimensional convolutional layer.

[0040] Optionally, the crisscross attention module includes at least two crisscross attention units connected in sequence, each crisscross attention unit including a dual-path crisscross attention layer and a two-dimensional convolutional layer;

[0041] In the first cross-attention unit, the embedded features of the temporal path and the embedded features of the spatial path are input into the dual-path cross-attention layer, which outputs the attention results of the temporal path and the attention results of the spatial path. The attention results of the temporal path and the attention results of the spatial path are integrated into a tensor through the outer product, and the tensor is input into the two-dimensional convolution layer, which outputs the two convolution results.

[0042] In the cross attention unit between the first cross attention unit and the last cross attention unit, the two convolution results output by the previous cross attention unit are input into the dual-path cross attention layer, which outputs two attention results. The two attention results are integrated into a tensor through the outer product, and the tensor is input into the two-dimensional convolution layer, which outputs two convolution results.

[0043] In the last cross-attention unit, the two convolution results output by the previous cross-attention unit are input into the dual-path cross-attention layer, which outputs two attention results. The two attention results are integrated into a tensor through the outer product, and the tensor is input into the two-dimensional convolution layer to output the classification label of the rs-fMRI data.

[0044] In a second aspect, the present invention provides an rs-fMRI data intelligent recognition device, comprising:

[0045] A preprocessing module is used to preprocess rs-fMRI data, which contains multiple time series;

[0046] A data recognition module is used to input the preprocessed rs-fMRI data into a pre-trained rs-fMRI recognition model to obtain the classification label of the rs-fMRI data;

[0047] The rs-fMRI recognition model includes a temporal feature extraction path, a spatial feature extraction path, and a cross-attention module. In the temporal feature extraction path, multiple time series are input into the long short-term memory (LSTM) network, and the hidden state sequence corresponding to each time series is output. The perturbation position code corresponding to each time series is generated according to the position index in the time series and the random perturbation noise. The enhanced features corresponding to each time series obtained by adding the hidden state sequence and the perturbation position code corresponding to the same time series are output as the embedded features of the temporal path. In the spatial feature extraction path, the Gramian angle field is used to convert each time series into a GAF image, and the structural similarity index between GAF images is calculated. The whole-brain functional connectivity map is constructed based on the structural similarity index between GAF images. The whole-brain functional connectivity map is input into the graph convolutional network, and the embedded features of the spatial path are output. The embedded features of the temporal path and the embedded features of the spatial path are input into the cross-attention module, and the classification label of the rs-fMRI data is output.

[0048] (3) Beneficial effects

[0049] The beneficial effects of the present invention are:

[0050] The intelligent rs-fMRI data recognition method and device provided in this embodiment utilize the excellent feature extraction and pattern recognition capabilities of a novel deep learning architecture to more effectively and accurately recognize rs-fMRI data. By introducing perturbation position encoding and spatial feature embedding based on the Gramian angular field similarity index, the model can better handle the physiological and behavioral differences between individuals and improve the model's adaptability to individual differences. Compared with traditional manual feature extraction and statistical analysis methods, the automated recognition method of this embodiment significantly improves data processing speed and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The present invention is described with the aid of the following drawings:

[0052] Figure 1 1 is a flow chart of the rs-fMRI data intelligent recognition method according to Example 1;

[0053] Figure 2 Schematic diagram of a process flow of inputting pre-processed rs-fMRI data into a pre-trained rs-fMRI recognition model according to Example 1;

[0054] Figure 3 2 is a schematic diagram of the processing flow of the first cross-attention unit according to Example 2;

[0055] Figure 4 1 is a schematic diagram of a processing flow of a cross-attention unit between the first cross-attention unit and the last cross-attention unit according to Example 2;

[0056] Figure 5 Schematic diagram of the processing flow of the last cross-attention unit according to Example 2. DETAILED DESCRIPTION

[0057] In order to better explain the present invention and facilitate understanding, the present invention is described in detail below through specific implementation methods in conjunction with the accompanying drawings.

[0058] Example 1

[0059] like Figure 1 and Figure 2 As shown, this embodiment provides an rs-fMRI data intelligent recognition method, comprising the following steps:

[0060] S1. Preprocessing rs-fMRI data. rs-fMRI data contains multiple time series.

[0061] Specifically, the rs-fMRI data were preprocessed, including:

[0062] S01. Spatial normalization of the rs-fMRI data. This allows all subjects' rs-fMRI data to be mapped into the same standardized anatomical space, ensuring that data analysis is performed within a uniform anatomical structure.

[0063] S02. Truncate multiple time series according to a pre-set sequence length, thereby ensuring consistent data analysis.

[0064] S03. Fill in missing values in multiple time series to enhance data integrity.

[0065] Furthermore, interpolation methods or machine learning models are used to fill missing values in multiple time series.

[0066] S2. Input the preprocessed rs-fMRI data into the pre-trained rs-fMRI recognition model to obtain the classification label of the rs-fMRI data.

[0067] The rs-fMRI recognition model includes a temporal feature extraction path, a spatial feature extraction path, and a cross-attention module. In the temporal feature extraction path, multiple time series are input into the long short-term memory network (LSTM), and the hidden state sequence corresponding to each time series is output. The perturbation position code corresponding to each time series is generated according to the position index in the time series and the random perturbation noise. The hidden state sequence and the perturbation position code corresponding to the same time series are added together to obtain the enhanced features corresponding to each time series, and the enhanced features corresponding to each time series are output as the embedded features of the temporal path. In the spatial feature extraction path, the Gramian angle field is used to convert each time series into a GAF image, and the structural similarity index between GAF images is calculated. The whole-brain functional connectivity map is constructed based on the structural similarity index between GAF images. The whole-brain functional connectivity map is input into the graph convolutional network, and the embedded features of the spatial path are output. The embedded features of the temporal path and the embedded features of the spatial path are input into the cross-attention module, and the classification label of the rs-fMRI data is output.

[0068] In the intelligent recognition method for rs-fMRI data provided in this embodiment, a novel deep learning architecture is designed. It comprises two parallel pathways—a temporal pathway and a spatial pathway—to process the temporal dynamics and spatial structural features of resting-state functional MRI (rs-fMRI) data, respectively. Specifically, an LSTM is used to analyze time series to extract temporal dynamic features, while a graph convolutional network is used to analyze functional connectivity between brain regions, capturing spatial structural features. To address signal inconsistencies caused by individual differences, this embodiment introduces perturbed position encoding in the temporal pathway, adding random noise to the position encoding to improve the model's adaptability to time series variations. Considering that rs-fMRI signals may exhibit instability in local dynamic relationships due to multicenter bias and individual differences, this embodiment uses the Gramian angle field to describe the fMRI sequence of each brain region in the spatial pathway to capture global signal characteristics and subtle variations. Furthermore, the structural similarity index (GAF) is calculated to assess the similarity between images. This index serves as a measure of functional connectivity between brain regions and more accurately captures the strength of functional connectivity between regions. The model of this embodiment can integrate spatiotemporal information more deeply by implementing cross-attention operations between temporal paths and spatial paths, thereby improving the expressiveness of features.

[0069] In summary, the intelligent rs-fMRI data recognition method provided in this embodiment, with the help of the excellent feature extraction and pattern recognition capabilities of the new deep learning architecture, can more effectively and accurately recognize rs-fMRI data; by introducing perturbation position encoding and spatial feature embedding based on the Gramian angular field similarity index, the model can better handle the physiological and behavioral differences between different individuals and improve the model's adaptability to individual differences; compared with traditional manual feature extraction and statistical analysis methods, the automated recognition method of this embodiment significantly improves data processing speed and efficiency.

[0070] Specifically, the data processing process of the time feature extraction path is expressed as follows:

[0071]

[0072] Where, is the embedding feature of the time path, R is the size of PPE, m is the time point, d e is the dimension of the embedded features; ε T is an LSTM network; X is rs-fMRI data; PPE is the perturbation position encoding.

[0073] Specifically, the disturbance position code corresponding to each time series is generated according to the position index in the time series and the random disturbance noise, including:

[0074]

[0075] Where PPE is the perturbation position code corresponding to the time series; pos is the position index in the time series; p is the random disturbance noise; d e is the dimension of the embedded feature; i is the count of the time point; i%2 is i divisible by 2; dim is the dimension; is the size of PPE, and m is the time point.

[0076] Specifically, each time series is converted into a GAF image using the Gramian angle field, which is expressed as follows:

[0077]

[0078] Where, X c is the time series of the cth region of interest (region of interest refers to the brain area) of rs-fMRI; c For X c GAF image; I is the identity matrix; T is the transpose operation of the matrix.

[0079] Specifically, the whole-brain functional connectivity map is constructed based on the structural similarity index between GAF images and GAF images. The formula is expressed as:

[0080]

[0081] Where, is the structural similarity index between GAF images; and They are the GAF images of the c1 channel and the GAF images of the c2 channel respectively; and They are and The average value of and They are and The standard deviation of for and covariance of ; δ1 and δ2 are both constants.

[0082] Specifically, the spatial feature extraction path also includes: performing sparse processing on the whole-brain functional connectivity map according to a pre-set threshold to obtain an adjacency matrix, constructing a sparse whole-brain functional connectivity map based on the adjacency matrix, inputting the sparse whole-brain functional connectivity map into a graph convolutional network, and outputting the embedded features of the spatial path;

[0083] The formula of the adjacency matrix is expressed as:

[0084]

[0085] Where, is the adjacency matrix; is the structural similarity index between GAF images; τ is a pre-set threshold.

[0086] The sparse whole-brain functional connectivity graph is represented as G = {V, E}, where V is a node set, each node represents a region of interest, and E is an edge set, representing the functional connectivity of the region of interest.

[0087] Specifically, the sparse whole-brain functional connectivity map is input into the graph convolutional network, and the embedded features of the output spatial path are expressed as follows:

[0088]

[0089] Where, is the embedding feature of the spatial path, ε s It is a graph convolutional neural network.

[0090] Preferably, the cross-attention module includes a dual-path cross-attention layer and a two-dimensional convolutional layer; the embedded features of the temporal path and the embedded features of the spatial path are input into the dual-path cross-attention layer, the attention results of the temporal path and the attention results of the spatial path are output, the attention results of the temporal path and the attention results of the spatial path are integrated into a tensor through the outer product, the tensor is input into the two-dimensional convolutional layer, and the classification label of the rs-fMRI data is output. In this way, in the dual-path cross-attention layer, by exchanging the key vectors of the temporal path and the spatial path, a cross-path attention mechanism is implemented to promote the deep fusion of spatiotemporal information; integrating the dual-path features into a single tensor can avoid the loss and deformation of spatiotemporal information and preserve its spatiotemporal dimensions; then, feed-forward processing is performed using two-dimensional convolution to enhance spatiotemporal correlation analysis at the local scale.

[0091] Specifically, the attention results of the temporal path and the attention results of the spatial path are integrated into a tensor through the outer product, which is expressed as follows:

[0092] P = ρ S ×ρ T

[0093] Where P is a tensor, ρ S is the attention result of the time path; ρ T is the attention result of the spatial path.

[0094] Specifically, the tensor is input into the two-dimensional convolutional layer, and the classification label of the rs-fMRI data is output. The formula is expressed as:

[0095] Z=Conv 2D (P)

[0096] Where Z is the classification label of rs-fMRI data, Conv 2D It is a two-dimensional convolutional layer.

[0097] It should be noted that before S1, there is also a training process, which is basically the same as S1 and S2 and will not be repeated here. During the training process, supervised learning is performed by minimizing the cross entropy loss function and combining the real classification labels.

[0098] Example 2

[0099] The rs-fMRI data intelligent recognition method provided in this embodiment differs from that in Example 1 in that:

[0100] The crisscross attention module includes at least two crisscross attention units connected in sequence, each of which includes a dual-path crisscross attention layer and a two-dimensional convolutional layer.

[0101] like Figure 3 As shown, in the first cross-attention unit, the embedded features of the temporal path and the embedded features of the spatial path are input into the dual-path cross-attention layer, which outputs the attention results of the temporal path and the attention results of the spatial path. The attention results of the temporal path and the attention results of the spatial path are integrated into a tensor through the outer product, and the tensor is input into the two-dimensional convolution layer, which outputs two convolution results.

[0102] If there is a cross-attention unit between the first cross-attention unit and the last cross-attention unit, then Figure 4 As shown in the figure, in the cross-attention unit, the two convolution results output by the previous cross-attention unit are input into the dual-path cross-attention layer, which outputs two attention results. The two attention results are integrated into a tensor through the outer product, and the tensor is input into the two-dimensional convolution layer, which outputs two convolution results.

[0103] like Figure 5 As shown in the figure, in the last cross attention unit, the two convolution results output by the previous cross attention unit are input into the dual-path cross attention layer, which outputs two attention results. The two attention results are integrated into a tensor through the outer product, and the tensor is input into the two-dimensional convolution layer to output the classification label of the rs-fMRI data.

[0104] The rest of the contents are the same as those in Example 1 and will not be repeated here.

[0105] Example 3

[0106] This embodiment provides an rs-fMRI data intelligent recognition device, comprising:

[0107] The preprocessing module is used to preprocess the rs-fMRI data, which contains multiple time series.

[0108] The data recognition module is used to input the preprocessed rs-fMRI data into the pre-trained rs-fMRI recognition model to obtain the classification label of the rs-fMRI data.

[0109] The rs-fMRI recognition model includes a temporal feature extraction path, a spatial feature extraction path, and a cross-attention module. In the temporal feature extraction path, multiple time series are input into the long short-term memory (LSTM) network, and the hidden state sequence corresponding to each time series is output. The perturbation position code corresponding to each time series is generated according to the position index in the time series and the random perturbation noise. The enhanced features corresponding to each time series obtained by adding the hidden state sequence and the perturbation position code corresponding to the same time series are output as the embedded features of the temporal path. In the spatial feature extraction path, the Gramian angle field is used to convert each time series into a GAF image, and the structural similarity index between GAF images is calculated. The whole-brain functional connectivity map is constructed based on the structural similarity index between GAF images. The whole-brain functional connectivity map is input into the graph convolutional network, and the embedded features of the spatial path are output. The embedded features of the temporal path and the embedded features of the spatial path are input into the cross-attention module, and the classification label of the rs-fMRI data is output.

[0110] It should be noted that the specific functions of each module in the rs-fMRI data intelligent recognition device provided in this embodiment, as well as the rs-fMRI data intelligent recognition process, can be referred to the detailed description of the rs-fMRI data intelligent recognition method provided in the above-mentioned Example 1 or Example 2, and will not be repeated here.

[0111] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0112] The present invention is described with reference to flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions.

[0113] It should be noted that, in the claims, any reference signs placed between brackets shall not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention may be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In claims enumerating several means, several of these means may be embodied by one and the same hardware. The use of the words first, second, third etc. is for convenience only and does not indicate any order. These words may be understood as part of the component name.

[0114] In addition, it should be noted that, in the description of this specification, the description of the terms "one embodiment", "some embodiments", "embodiment", "example", "specific example" or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and the features of different embodiments or examples, unless they are contradictory.

[0115] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments after learning the basic creative concept. Therefore, the claims should be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0116] Obviously, those skilled in the art may make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if such modifications and variations fall within the scope of the claims and their equivalents, the present invention shall also include such modifications and variations.

Claims

1. A method for intelligent recognition of rs-fMRI data, characterized in that: The following steps are involved: S1. Preprocessing rs-fMRI data. rs-fMRI data contains multiple time series. S2, input the pre-processed rs-fMRI data into the pre-trained rs-fMRI recognition model to obtain the classification label of the rs-fMRI data; The rs-fMRI recognition model includes a temporal feature extraction path, a spatial feature extraction path, and a cross-attention module. In the temporal feature extraction path, multiple time series are input into a long short-term memory (LSTM) network, and the hidden state sequence corresponding to each time series is output. The perturbation position code corresponding to each time series is generated based on the position index in the time series and random perturbation noise. The enhanced features corresponding to each time series obtained by adding the hidden state sequence and perturbation position code corresponding to the same time series are output as the embedded features of the temporal path. In the spatial feature extraction path, the Gramian angle field is used to convert each time series into a GAF image, and the structural similarity index between GAF images is calculated. The whole-brain functional connectivity map is constructed based on the structural similarity index between GAF images. The whole-brain functional connectivity map is input into a graph convolutional network, and the embedded features of the spatial path are output. The embedded features of the temporal path and the spatial path are input into the cross-attention module to output the classification label of the rs-fMRI data.

2. The rs-fMRI data intelligent recognition method according to claim 1, characterized in that: Preprocessing of rs-fMRI data includes: S01, perform spatial normalization on rs-fMRI data; S02. truncating multiple time series according to a preset sequence length; S03. Fill in missing values in multiple time series.

3. The rs-fMRI data intelligent recognition method according to claim 1 or 2, characterized in that: Generate the disturbance position code corresponding to each time series according to the position index and random disturbance noise in the time series, including: Where PPE is the perturbation position code corresponding to the time series; pos is the position index in the time series; p is the random disturbance noise; d e is the dimension of the embedded feature; i is the count of the time point; i%2 is i divisible by 2; dim is the dimension; is the size of PPE, and m is the time point.

4. The rs-fMRI data intelligent recognition method according to claim 1, characterized in that: Each time series is converted into a GAF image using the Gramian angle field. The formula is: Where, X c is the time series of the cth region of interest of rs-fMRI; χ c For X c GAF image; I is the identity matrix; T is the transpose operation of the matrix.

5. The rs-fMRI data intelligent recognition method according to claim 1, characterized in that: The whole-brain functional connectivity map is constructed based on the structural similarity index between GAF images and GAF images. The formula is expressed as: Where, is the structural similarity index between GAF images; and They are the GAF images of the c1 channel and the GAF images of the c2 channel respectively; and They are and The average value of and They are and The standard deviation of for and covariance of ; δ1 and δ2 are both constants.

6. The rs-fMRI data intelligent recognition method according to claim 5, characterized in that: The spatial feature extraction path also includes: performing sparse processing on the whole-brain functional connectivity map according to a pre-set threshold to obtain an adjacency matrix, constructing a sparse whole-brain functional connectivity map based on the adjacency matrix, inputting the sparse whole-brain functional connectivity map into a graph convolutional network, and outputting embedded features of the spatial path; The formula of the adjacency matrix is expressed as: Where, is the adjacency matrix; is the structural similarity index between GAF images; τ is a pre-set threshold.

7. The rs-fMRI data intelligent recognition method according to claim 1, characterized in that: The crisscross attention module includes a dual-path crisscross attention layer and a two-dimensional convolutional layer; The embedded features of the temporal path and the embedded features of the spatial path are input into the dual-path cross attention layer, and the attention results of the temporal path and the attention results of the spatial path are output. The attention results of the temporal path and the attention results of the spatial path are integrated into a tensor through the outer product, and the tensor is input into the two-dimensional convolutional layer to output the classification label of the rs-fMRI data.

8. The rs-fMRI data intelligent recognition method according to claim 7, characterized in that: The attention results of the time path and the attention results of the spatial path are integrated into a tensor through the outer product, and the formula is expressed as: P=p S ×r T Where P is a tensor, R is the size of PPE, m is the time point, d e is the dimension of the embedded feature; ρ S is the attention result of the time path; ρ T is the attention result of the spatial path; The tensor is input into the two-dimensional convolutional layer, and the classification label of the rs-fMRI data is output. The formula is expressed as: Z=Conv 2D (P) Where Z is the classification label of rs-fMRI data, R is the size of PPE, m is the time point, d e is the dimension of the embedded feature; Conv 2D It is a two-dimensional convolutional layer.

9. The rs-fMRI data intelligent recognition method according to claim 1, characterized in that: The crisscross attention module includes at least two crisscross attention units connected in sequence, each crisscross attention unit includes a dual-path crisscross attention layer and a two-dimensional convolutional layer; In the first cross-attention unit, the embedded features of the temporal path and the embedded features of the spatial path are input into the dual-path cross-attention layer, which outputs the attention results of the temporal path and the attention results of the spatial path. The attention results of the temporal path and the attention results of the spatial path are integrated into a tensor through the outer product, and the tensor is input into the two-dimensional convolution layer, which outputs the two convolution results. In the cross attention unit between the first cross attention unit and the last cross attention unit, the two convolution results output by the previous cross attention unit are input into the dual-path cross attention layer, which outputs two attention results. The two attention results are integrated into a tensor through the outer product, and the tensor is input into the two-dimensional convolution layer, which outputs two convolution results. In the last cross-attention unit, the two convolution results output by the previous cross-attention unit are input into the dual-path cross-attention layer, which outputs two attention results. The two attention results are integrated into a tensor through the outer product, and the tensor is input into the two-dimensional convolution layer to output the classification label of the rs-fMRI data.

10. An intelligent recognition device for rs-fMRI data, characterized in that: include: A preprocessing module is used to preprocess rs-fMRI data, which contains multiple time series; A data recognition module is used to input the preprocessed rs-fMRI data into a pre-trained rs-fMRI recognition model to obtain the classification label of the rs-fMRI data; The rs-fMRI recognition model includes a temporal feature extraction path, a spatial feature extraction path, and a cross-attention module. In the temporal feature extraction path, multiple time series are input into a long short-term memory (LSTM) network, and the hidden state sequence corresponding to each time series is output. The perturbation position code corresponding to each time series is generated based on the position index in the time series and random perturbation noise. The enhanced features corresponding to each time series obtained by adding the hidden state sequence and perturbation position code corresponding to the same time series are output as the embedded features of the temporal path. In the spatial feature extraction path, the Gramian angle field is used to convert each time series into a GAF image, and the structural similarity index between GAF images is calculated. The whole-brain functional connectivity map is constructed based on the structural similarity index between GAF images. The whole-brain functional connectivity map is input into a graph convolutional network, and the embedded features of the spatial path are output. The embedded features of the temporal path and the spatial path are input into the cross-attention module to output the classification label of the rs-fMRI data.

Images