Adhd diagnostic system and method fusing neurodynamic enhancement and domain adaptation

Abstract

A diagnostic system and method for ADHD integrating neurodynamic enhancement and domain adaptation are proposed. This belongs to the interdisciplinary field of medical image processing and artificial intelligence. It solves the technical problems of low accuracy, poor generalization ability, and poor robustness and interpretability in existing ADHD diagnostic methods. The system includes a data input and preprocessing module, a neuromass-chaotic perturbation module, a domain adaptation module with mutual information minimization, and a multimodal graph convolutional network module. The neuromass-chaotic perturbation module captures the nonlinear dynamic characteristics of brain activity, the domain adaptation module effectively eliminates domain differences in multi-site data, graph structure constraint loss ensures the consistency of sample association, and the dynamic characteristics can correspond to the pathological mechanisms of brain functional abnormalities, providing interpretable evidence for clinical diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing and artificial intelligence interdisciplinary technology, specifically involving an ADHD diagnostic system and method that integrates neurodynamic enhancement and domain adaptation. Background Technology

[0002] Existing ADHD identification technologies based on functional magnetic resonance imaging (fMRI) have the following three main drawbacks:

[0003] 1. Ignoring the nonlinear dynamics of brain activity: Traditional methods typically use raw fMRI time series or extract simple statistical features such as mean and standard deviation as model inputs. However, brain activity is inherently a complex nonlinear dynamic system with characteristics such as chaos and bifurcation. Existing methods fail to effectively model and utilize this deep-seated dynamic information, resulting in insufficient feature representation capabilities and limiting the improvement of recognition accuracy.

[0004] 2. Poor generalization ability across sites: In clinical studies, fMRI data is typically collected from multiple sites (hospitals, research centers). Due to differences in scanning equipment models, scanning parameters, and subject conditions, significant "domain shift" occurs between data from different sites. Models trained on data from a single site will perform significantly worse on data from other sites, lacking generalization ability and hindering the clinical deployment of the technology.

[0005] 3. Insufficient Feature Fusion and Association Modeling: Existing methods, when constructing brain region association maps, often rely on single functional connectivity (such as the Pearson correlation coefficient) or simply fuse multiple features, failing to dynamically and adaptively balance the importance of different association patterns (such as feature similarity and functional connectivity). Furthermore, they lack sufficient modeling of the overall similarity relationships among subjects (samples), neglecting group-level association information. Summary of the Invention

[0006] To address at least one of the technical problems existing in the aforementioned background art, a first aspect of the present invention provides an ADHD diagnostic system integrating neurodynamic enhancement and domain adaptation, the system comprising:

[0007] Data Input and Preprocessing Module: Receives ROI time-series data from multiple sites based on functional magnetic resonance imaging (fMRI) data from subjects. The sample similarity matrix is ​​obtained through basic feature construction, adaptive affinity fusion, and sample similarity matrix construction. ;

[0008] The neural quality-chaotic perturbation module uses a neural quality model to perform basic brain region dynamics simulations on several ROI time series sets, obtaining the simulated data. A nonlinear perturbation signal is generated using a Rössler chaotic system; the nonlinear perturbation signal is then injected... Perform chaotic dynamics fusion to obtain enhanced time series ;from Eight high-discrimination features were extracted to form a dynamic feature matrix. ;

[0009] Domain adaptation module with minimized mutual information: raw image feature matrix of functional magnetic resonance imaging data Based on logistic regression, the model iteratively removes features with the lowest importance until the feature dimension is reached. Stop and output the feature matrix after iteration. ;

[0010] Multimodal Graph Convolutional Network Module: By constructing a graph structure... and Perform node feature fusion to obtain the node feature matrix. And construct the adjacency matrix It reflects the strength of the association between nodes; it captures the relationship through graph convolution. Brain region association features in the middle, output compressed association features ; By classifying Mapping to the classification dimension yields the classification feature matrix. ,pass right Perform probability distribution transformation to output the probability that the subject is an ADHD patient.

[0011] Furthermore, time-series ROI data from multiple sites based on functional magnetic resonance imaging (fMRI) data of subjects. ,in For the sample size, For time points, Enter the number of brain regions and their corresponding task tags. and domain tags , For healthy controls, For patients with ADHD, domain labels are used to mark collection sites.

[0012] Furthermore, the sample similarity matrix is ​​obtained as follows:

[0013] S31. Construction of Basic Features: From Extract the mean and standard deviation of each ROI to obtain the first... The first sample Time series mean of each ROI , and the The first sample Time series standard deviation of ROI ; and Together constitute ;

[0014] S32, Adaptive Affinity Fusion: Based on Gaussian kernel similarity is used to construct the feature similarity matrix. , ;based on Pearson correlation coefficients are used to construct functional connectivity matrices. ;Will and Perform fusion to obtain fusion characteristics ;

[0015] S33. Construction of the similarity matrix between samples:

[0016] S331, Sample-level affinity graph extension: Calculate the affinity matrix for each sample separately, and obtain... :

[0017] ;

[0018] S332, Sample-level statistical feature extraction: Take the mean of all ROI features for each sample to obtain... :

[0019] ;

[0020] in For the sample The aggregated mean of all ROIs For the sample Aggregate standard deviation of all ROIs;

[0021] S333, Sample Similarity Matrix: The sample similarity matrix is ​​calculated using Gaussian kernel similarity. value range , Indicates sample With sample The similarity.

[0022] Furthermore, the eight high-discrimination features include: the maximum Lyapunov index. Sample entropy Hurst index Low, medium and high frequency energy , and Disturbance variance and local synchronization ; Dynamic characteristic matrix .

[0023] Furthermore, in the domain adaptation module that minimizes mutual information, a mutual information definition operation is also performed for features. ,in, Representation of features Feature dimension number, domain label and task tags Define the mutual information of features and domain labels respectively. And mutual information between features and task labels :

[0024] ;

[0025] ;

[0026] in, This represents the total number of all sites participating in the data collection; express Take a specific feature value and domain tags The probability of taking a specific value at a particular site. express Take a specific feature value The probability of that time. Indicates field label Take a specific site value The probability of that time; express Take a specific feature value And task tags The probability of taking a specific probability value. Indicates task tags Take a specific probability value The probability of that time.

[0027] Furthermore, in multimodal graph convolutional networks, node feature fusion is specifically performed as follows:

[0028] Each node represents a subject sample, and the node features are image features. With dynamic characteristics: The integration;

[0029] First, perform the aggregation operation: for each subject sample... The global features are obtained by averaging the features of all ROIs.

[0030] Global image features ;

[0031] Global dynamics ;

[0032] Then, feature fusion is performed, stitching together the global image features and global dynamic features of the subject samples according to their dimensions to obtain the node feature matrix:

[0033] ;

[0034] in For the first The feature vectors of each node have the same dimensions as the selected features in the document.

[0035] In multimodal graph convolutional networks, the adjacency matrix is ​​constructed. Specifically:

[0036] Calculate the degree matrix: This calculates the total edge weights for each node, reflecting the global interconnectivity of the nodes.

[0037] ;

[0038] ;

[0039] It is the identity matrix;

[0040] in, Degree matrix The Middle The sum of the total edge weights of all nodes;

[0041] Normalized adjacency matrix: Eliminates the impact of node degree differences on convolution, ensuring balanced feature aggregation weights for different subjects.

[0042] ;

[0043] Output the final adjacency matrix ,in For nodes and Normalized edge weights.

[0044] Furthermore, capturing graph convolutions Brain region association features in the middle, output compressed association features Specifically:

[0045] Layer 1 graph convolution extracts primary association features: Input and By combining inter-subject similarity and subject characteristics through convolutional computation, primary association features are extracted. :

[0046] ;

[0047] in These are learnable weights;

[0048] The second layer of graph convolution deepens the associated features: inputting the features from the first layer of convolution. and Further deep connections between participants are explored through convolutional computation, feature dimensions are compressed to reduce complexity, and compressed connection features are output. :

[0049] ;

[0050] in These are learnable weights.

[0051] Furthermore, the classifier uses graph pooling layers and fully connected layers to... Mapping to the classification dimension to obtain the classification feature matrix ,pass Output prediction probability , This represents the probability that the sample is a healthy control. This represents the probability that the sample consists of ADHD patients.

[0052] A second aspect of the present invention provides a diagnostic method for ADHD that integrates neurodynamic enhancement and domain adaptation, the method comprising:

[0053] S1. Collect data from subjects at multiple sites based on functional magnetic resonance imaging and divide the data into training and validation sets;

[0054] S2. Input the training set data into the ADHD diagnostic system that integrates neurodynamic enhancement and domain adaptation, perform end-to-end training using the total loss function, and evaluate the system's predictive performance using the validation set.

[0055] S3. Apply the qualified system to the diagnosis of the probability of a subject having ADHD.

[0056] Furthermore, the total loss function is:

[0057] ;in, Cross-entropy classification loss measures the difference between the predicted ADHD probability and the true label. Graph structure constraint loss measures the difference between the predicted sample similarity matrix and the true sample similarity matrix. To reduce the association between features and domain labels, the domain adaptation loss is used. and They are respectively and The weighting coefficients.

[0058] The beneficial effects of the system described in this invention are as follows:

[0059] High recognition accuracy: By capturing the nonlinear dynamic characteristics of brain activity through the neural quality-chaotic perturbation module, multimodal feature fusion comprehensively describes brain region abnormalities, and the classification accuracy is improved by 5%-10% compared with existing methods;

[0060] Strong generalization ability: The domain adaptation module effectively eliminates domain differences in multi-site data, improving cross-site recognition accuracy by 8%-15%, meeting the needs of clinical multi-center applications;

[0061] Good robustness and interpretability: Graph structure constraint loss ensures the consistency of sample association, and dynamic features (such as Lyapunov index and sample entropy) can correspond to the pathological mechanism of brain dysfunction, providing interpretable evidence for clinical diagnosis;

[0062] Highly feasible for engineering implementation: All modules are implemented based on mature mathematical models and deep learning frameworks (such as PyTorch and TensorFlow), with controllable computational complexity, and can be deployed in clinical auxiliary diagnostic systems. Attached Figure Description

[0063] Figure 1 This is a flowchart illustrating the workflow of the ADHD diagnostic system that integrates neurodynamic enhancement and domain adaptation in an embodiment of the present invention. Detailed Implementation

[0064] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0065] Example 1

[0066] This embodiment provides an ADHD diagnostic system that integrates neurodynamic enhancement and domain adaptation, including:

[0067] Data Input and Preprocessing Module: Receives ROI time-series data from multiple sites based on functional magnetic resonance imaging (fMRI) data from subjects. The sample similarity matrix is ​​obtained through basic feature construction, adaptive affinity fusion, and sample similarity matrix construction. ;

[0068] The neural quality-chaotic perturbation module uses a neural quality model to perform basic brain region dynamics simulations on several ROI time series sets, obtaining the simulated data. A nonlinear perturbation signal is generated using a Rössler chaotic system; the nonlinear perturbation signal is then injected... Perform chaotic dynamics fusion to obtain enhanced time series ;from Eight high-discrimination features were extracted to form a dynamic feature matrix. ;

[0069] Domain adaptation module with minimized mutual information: raw image feature matrix of functional magnetic resonance imaging data Based on logistic regression, the model iteratively removes features with the lowest importance until the feature dimension is reached. Stop and output the feature matrix after iteration. ;

[0070] Multimodal Graph Convolutional Network Module: By constructing a graph structure... and Perform node feature fusion to obtain the node feature matrix. And construct the adjacency matrix It reflects the strength of the association between nodes; it captures the relationship through graph convolution. Brain region association features in the middle, output compressed association features ; By classifying Mapping to the classification dimension yields the classification feature matrix. ,pass right Perform probability distribution transformation to output the probability that the subject is an ADHD patient.

[0071] The core of this embodiment is to propose a "Chaotic Adaptive Multi-modal GCN" (CAM-GCN), which enhances nonlinear dynamic features through a neural mass model (NMM) and chaotic perturbation, constructs robust brain region maps using adaptive affinity fusion, and achieves domain adaptation by minimizing mutual information, ultimately improving the accuracy and robustness of ADHD cross-site identification.

[0072] Example 2

[0073] This embodiment further defines Embodiment 1 and provides further explanation of the data input and preprocessing module.

[0074] The specific operations performed in the data input and preprocessing module are as follows:

[0075] 1. fMRI ROI time series input

[0076] Input multi-site fMRI ROI time series data The BOLD signal of each brain region was preprocessed, among which... For the sample size, For time points (truncation is uniform). Input the number of brain regions; simultaneously input the corresponding task tags. ( For healthy controls, (for ADHD patients) and domain labels (Mark the data collection site).

[0077] 2. Construction of basic features

[0078] ROI statistical characteristics: from Extract the mean and standard deviation of each ROI to obtain The basic set of statistical features extracted from time series data (first dimension) The second dimension is the number of samples. The third dimension, 2, represents the number of regions of interest (ROIs). It consists of two fixed core statistics: one is... Corresponding to the The first sample The time series mean of one ROI, and the other is... Corresponding to the The first sample The time series standard deviation of each ROI is used to construct a brain region affinity map. The specific calculation is as follows:

[0079] ;

[0080] ;

[0081] in, For sample index, For ROI indexing, For the nth sample Time series mean of each ROI For the first The first sample The time series standard deviation of each ROI.

[0082] 3. Adaptive Affinity Fusion

[0083] To fully describe the association strength between brain regions, relying solely on ROI time-series features (without phenotypic information), the following structure is constructed:

[0084] (1) Feature similarity matrix: based on Gaussian kernel similarity:

[0085] ;

[0086] in For adaptive bandwidth, .

[0087] This represents the square of the L2 norm (Euclidean norm). This represents the arithmetic mean operation function, which sums all elements in a specified dataset and then divides by the total number of elements to obtain the average value of the dataset. The subscript 2 in the symbol represents the L2 norm (Euclidean norm), a commonly used mathematical operator for measuring the length of a vector or the distance between two vectors.

[0088] (2) Functional connection matrix: based on Pearson correlation coefficient:

[0089] ;

[0090] in For the first Time series mean of each ROI.

[0091] (3) Fusion characteristics: Balancing two association patterns:

[0092] ;

[0093] 4. Construction of the similarity matrix between samples

[0094] A region-of-origin affinity map is obtained by fusing the feature similarity matrix and the functional connectivity matrix to describe the strength of associations between brain regions. To further capture the overall feature associations between samples (subjects), a "sample similarity matrix" is introduced to extract the similarity relationships between samples:

[0095] (1) Sample-level affinity graph expansion: Calculate the affinity matrix for each sample separately to obtain :

[0096] .

[0097] (2) Sample-level statistical feature extraction: Take the mean of all ROI features for each sample to obtain :

[0098] ;

[0099] in For the sample The aggregated mean of all ROIs This represents the aggregate standard deviation.

[0100] (3) Sample similarity matrix: calculated using Gaussian kernel similarity The value range is [0,1]. This represents the similarity between samples n and m.

[0101] Example 3

[0102] This embodiment further defines Embodiment 1, providing a more detailed explanation of the Neural Mass-Chaotic Perturbation Module (NM-CDM). Existing methods often directly use raw fMRI time series or simple statistical features, neglecting the nonlinear dynamic characteristics of brain activity. The NM-CDM module simulates the basic dynamics of brain regions through a neural mass model (NMM) and injects chaotic perturbations to enhance nonlinear features. It is the first method to combine a neural mass model with chaotic perturbations to model the nonlinear dynamic characteristics of brain activity, extracting eight highly discriminative dynamic features and overcoming the deficiency of existing methods in neglecting the nonlinear characteristics of the brain.

[0103] Specifically, it includes four key steps:

[0104] 1. Neural mass model (NMM) simulates the basic dynamics of brain regions:

[0105] The NMM-based neuronal population average activity equation is constructed with reference to a simplified version of the Jansen-Rit model. It describes the evolution of brain region membrane potential-related states and is closer to the physiological mechanisms of real brain activity.

[0106] For 7 ROI groups Model the state equations separately:

[0107] ;

[0108] in, For variables related to the average membrane potential of the neuronal population; The rate of change of membrane potential;

[0109] Connection strength, (Time constant);

[0110] Indicates the first The time series mean of the group's ROI serves as an external driver to ensure consistency with the original data. This represents the number of ROIs within the group.

[0111] Here is the Sigmoid activation function, where , ,

[0112] This simulates the nonlinear characteristics of neuronal firing.

[0113] Numerical solution: Euler method is used, with integration step size... Simulation duration To obtain the NMM output (The number of time points is aligned with the dynamic range of the original data).

[0114] These are the core data output by the Neural Quality Model (NMM) after simulating the physiological mechanisms of brain region activity:

[0115] First Dimension : Corresponding to the input sample size (i.e. the number of subjects, including healthy controls and ADHD patients), ensuring that the simulation results correspond one-to-one with the original samples;

[0116] Second dimension The corresponding time points are numerically solved using the Euler method (integration step size). Simulation duration This achieves temporal dynamic range alignment with the original fMRI data, ensuring consistency in the temporal dimension;

[0117] Third dimension : Corresponds to the brain region neuronal population activity state dimensions simulated by the neuron quality model (such as membrane potential-related states, firing frequency-related characteristics, etc.), used to characterize the multidimensional physiological characteristics of brain region dynamics;

[0118] Fourth Dimension The number of regions of interest (ROIs) corresponding to the brain regions ensures that the dynamic characteristics of each ROI can be simulated and output individually.

[0119] 2. Rössler chaotic system, generating nonlinear perturbation signals:

[0120] Brain activity inherently contains random noise and nonlinear fluctuations, which are difficult to model using traditional methods. Therefore, the Rössler chaotic system is introduced to generate perturbation signals with complex dynamic characteristics.

[0121] Chaos equations:

[0122] ;

[0123] in, , , For fixed parameters (ensuring the system remains in a chaotic state), initial conditions , , The solution yields a 3D chaotic signal. The number of time points is strictly aligned with the NMM output.

[0124] The state variables in the first dimension of a chaotic system, whose time evolution trajectory is the specific manifestation of the chaotic signal in the first dimension, are used to provide one of the fundamental components of chaotic disturbances.

[0125] The state variables of the second dimension of a chaotic system, and Through the coupling of differential equations, its evolution process introduces cross-dimensional dynamic correlations into chaotic signals, enhancing the complexity of the perturbation.

[0126] The third-dimensional state variable of a chaotic system, through interaction with... The nonlinear coupling triggers the chaotic behavior of the system, which is a key dimension for generating complex nonlinear disturbances. These three state variables, after being solved using differential equations, yield a 3D chaotic signal. Subsequently, ROI chaotic component mapping (evenly distributing 116 ROIs to three dimensions) will be used as nonlinear perturbation sources for different ROIs and injected into the output of the neural mass model (NMM) to enhance the characterization of the nonlinear dynamics of brain activity.

[0127] 3. Chaotic Dynamics Fusion Mechanism

[0128] Injecting chaotic signals into the NMM output preserves the fundamental dynamics while enhancing nonlinear characteristics. The steps are as follows:

[0129] (1) ROI chaotic component mapping: The 116 ROIs are evenly distributed to 3 chaotic components through modular operation (to ensure the perturbation is targeted):

[0130] No. Chaotic component indexes corresponding to each ROI: (The results are 0, 1, and 2 respectively) (Three dimensions).

[0131] Modulo operation (also known as remainder operation) is a basic mathematical operation. Its core logic is to calculate the remainder after dividing two integers, and the value of the remainder is always between "0" and "divisor - 1" (with the same sign as the divisor; here the divisor is 3, so the remainder is only 0, 1, or 2).

[0132] (2) Normalization of chaotic signals: Scale to the dynamic range of the NMM output to avoid excessively large or small disturbance amplitudes:

[0133] ;

[0134] in, It represents the values ​​of the original 3D chaotic signal generated by the Rössler chaotic system at a specific time point and at a specific component. For time point indexing (strictly aligned with the number of time points output by NMM; the number of time points for chaotic signals is 20000, representing any time point); Indexing the chaotic components (corresponding to the three dimensions of the chaotic signal), for example... This represents the original value of the chaotic signal at the 100th time point and the 2nd chaotic component.

[0135] For the original chaotic signal The result after normalization:

[0136] The purpose of normalization is to scale the numerical range of chaotic signals to the NMM output. The dynamic range should be set to avoid excessively large disturbances that could damage the fundamental dynamic characteristics of NMM, or too small disturbances that would fail to enhance nonlinearity.

[0137] This refers to the global maximum / minimum values ​​output by NMM.

[0138] (3) Element-by-element fusion: Controlling the perturbation intensity to generate enhanced time series:

[0139] ;

[0140] in To balance the fundamental dynamics and nonlinear disturbances for the disturbance intensity parameters, the final result is... .

[0141] 4. Dynamic Feature Extraction

[0142] from Eight high-discrimination features were extracted (reflecting the chaos, complexity, and rhythmicity of brain activity):

[0143] (1) Maximum Lyapunov index ( This is used to measure the degree of chaos. The system is considered chaotic, calculated using a small amount of data:

[0144] ;

[0145] in The distance to the nearest orbit. The time delay is determined based on the autocorrelation function.

[0146] (2) Sample entropy (SampEn) is used to measure the complexity of a sequence; the larger the value, the weaker the regularity.

[0147] ;

[0148] in For embedded dimensions, =0.2 (Similarity tolerance, (Standard deviation of the sequence) for The similarity number of pairs of vectors.

[0149] (3) Hurst exponent ( Long-range correlation is described and calculated using rescaled range analysis.

[0150] ;

[0151] in For recalibrated range, This represents the length of the time series.

[0152] (4-6): Low, medium and high frequency energy ( This reflects the rhythm of brain activity and is calculated through Fourier transform. (band)

[0153] low frequency ( ): ;

[0154] Intermediate frequency (IF) ): ;

[0155] high frequency( ): ;

[0156] (7) Disturbance variance ), measures the intensity of fluctuations in a time series:

[0157] ;

[0158] in .

[0159] (8) Local synchronization ( ): Calculate the phase consistency of ROIs within a group based on the synchronicity measure of Hilbert phase.

[0160] ;

[0161] in For the first Hilbert phase of each ROI.

[0162] Eight types of high-discrimination features (maximum Lyapunov exponent, sample entropy, Hurst exponent, low / medium / high frequency energy, perturbation variance, and local synchronization) are combined according to the "sample-brain region-feature" dimensional logic to form the final dynamic feature matrix. :

[0163] For each sample (N in total) and each brain region (R in total), the specific values ​​of the above 8 types of features are extracted to obtain the 8 feature values ​​corresponding to the "sample-brain region";

[0164] These 8 feature values ​​are concatenated according to their feature dimensions to form a feature vector with a dimension of 8;

[0165] Repeat the above operation for all samples and all brain regions, and finally stack them into a dimension of The matrix, i.e. .

[0166] The final dynamic characteristic matrix is: ;

[0167] in, This represents the total number of nonlinear dynamic features, where the meaning of the letter combinations is directly related to the extraction logic of dynamic features in the document.

[0168] Example 4

[0169] This embodiment further defines Embodiment 1 and provides a further explanation of the domain adaptation module with minimized mutual information.

[0170] Experimental data collection, especially with large datasets, reveals significant site (domain) differences (e.g., different scanning devices, parameters), leading to decreased model generalization ability. Mutual Information Minimization (MIM) is introduced to achieve domain adaptation, reducing the correlation between features and domain labels while preserving their correlation with task labels. Specifically:

[0171] 1. Definition of Mutual Information

[0172] Mutual information (MI) measures the dependency between two random variables, for features , It is the intermediate feature output by the Neurodynamic-Enhanced Disentangled GCN (NED-GCN) graph convolutional network, which is then output through the first layer of graph convolution. Z represents the number of feature dimensions of the intermediate features in a graph convolutional network, i.e., the number of features for each sample (subject) in the intermediate features. The dimension of the corresponding feature vector can be obtained through the design and parameter derivation of the graph convolutional layer.

[0173] For features Domain tags and task tags :

[0174] (1) Mutual information between features and domain labels:

[0175] ;

[0176] (2) Mutual information between features and task labels:

[0177] ;

[0178] in, This represents the total number of all sites participating in the data collection; express Take a specific feature value and domain tags The probability of taking a specific value at a particular site. express Take a specific feature value The probability of that time. Indicates field label Take a specific site value The probability of that time; express Take a specific feature value And task tags The probability of taking a specific probability value. Indicates task tags Take a specific probability value The probability of that time.

[0179] 2. Domain Adaptation Loss Function

[0180] By minimizing Simultaneously maximize This ensures that the features are "domain-invariant" and "task-relevant." The InfoNCE estimator is used to approximate mutual information, and the loss function is constructed as follows:

[0181] (1) Domain mutual information loss (minimizing and reducing the association between features and domains):

[0182] ;

[0183] in the formula It is the total number of samples (i.e., the number of subjects, corresponding to the feature matrix mentioned earlier). In The entire formula iterates through all... For each sample, calculate the loss term corresponding to each sample and then sum them up.

[0184] Molecular part The calculation method is as follows: find all samples... Samples belonging to the same domain For each sample Calculate samples With sample Feature similarity (dot product) divided by temperature coefficient The exponents are then summed.

[0185] The meaning of the variable is: Index of the sample (subjects) (and) correspond, This is the current sample. (This refers to the other samples being traversed). For the sample The domain label, i.e., the sample Collection sites / data sources; For the sample The domain label, i.e., the sample Collection sites / data sources;

[0186] condition Representation and Sample All samples with the same domain label .

[0187] Indicates: Sample intermediate feature vector transpose, It is a matrix The first in Rows, transposed into column vectors, used for Do the product.

[0188] Indicates: Sample intermediate feature vector (feature matrix) The OK)

[0189] Representation: Temperature coefficient, a hyperparameter used to control the weights of feature similarity. The smaller the similarity, the higher the distinguishability.

[0190] denominator The calculation method is as follows:

[0191] Traverse all samples Different samples Calculate each The corresponding feature similarity divided by The exponents are then summed.

[0192] It means: traversing all satisfying samples .

[0193] (2) Task mutual information loss (maximizing, preserving the association between features and tasks):

[0194] ;

[0195] in, Indicates sample The task tags correspond to the task objectives "ill / healthy"; Indicates sample The task tag corresponds to the task objective "ill / healthy".

[0196] (3) Total domain adaptation loss:

[0197] ;

[0198] in This is the balance coefficient.

[0199] Mutual information minimization domain adaptation: By minimizing the mutual information between features and domain labels and maximizing the mutual information between features and task labels, "domain invariant-task relevant" feature learning is achieved, which solves the problem of domain differences in multi-site data and improves the generalization of the model.

[0200] 3. Image Feature Selection (RFE)

[0201] By fusing image and dynamic features, a graph structure is constructed based on the similarity between samples. Two-layer graph convolution is used to capture deep correlations between samples, while feature filtering through RFE reduces redundancy and improves model efficiency and accuracy.

[0202] Because the original fMRI image features (such as ROI-based functional connectivity features, local consistency features, etc.) have high dimensionality and contain redundant information and noise, it is necessary to use recursive feature elimination (RFE) to select the features most relevant to the main task.

[0203] Input fMRI raw image feature matrix Based on logistic regression, the model iteratively removes features with the lowest importance until the feature dimension is reached. Stop. Output. As one of the subsequent inputs. The number of dimensions representing the features of the original fMRI image (i.e., the total number of original feature types).

[0204] Example 5

[0205] This embodiment further defines Embodiment 1 and provides a further explanation of the multimodal graph convolutional network module.

[0206] This module integrates image features and NM-CDM dynamic features, combined with a domain adaptation module, to achieve end-to-end classification. The specific processing flow is as follows:

[0207] 1. Construct a graph structure

[0208] (1) Node feature fusion

[0209] Each node represents a subject sample, and the node features are image features. (fMRI imaging features and dynamic features of RFE selected by feature selection): Fusion of (NM-CDM module output).

[0210] Aggregation operation: for each subject The global features are obtained by averaging the features of all ROIs.

[0211] ;

[0212] ;

[0213] Feature fusion: The global image and dynamic features of the subjects are stitched together according to their dimensions to obtain the node feature matrix.

[0214] ;

[0215] in For the first The feature vectors of each node (subject) have the same dimension as the dimension after feature selection.

[0216] (2) Edge construction and preprocessing

[0217] Through the sample similarity matrix Construct edges.

[0218] Preprocessing ensures graph convolution stability:

[0219] Adding self-loops: Preserves the node's own features and avoids losing individual subject information during convolution.

[0220] ;

[0221] This refers to the sample-level adjacency matrix after adding self-loops, which is derived from the sample similarity matrix. The result is obtained by adding the (description of the participants' association strength) to the identity matrix I (which provides self-loops), with dimensions of [missing information]. Its core function is to preserve the individual information of the subjects, avoid the loss of their own features during graph convolution, and lay the foundation for subsequent graph convolution calculations.

[0222] in It is the identity matrix. ,because .

[0223] Calculate the degree matrix: Calculate the total edge weights of each node (subject) to reflect the global interconnectivity of the nodes, thus obtaining the degree matrix. degree matrix In the middle, the first The total edge weights of each node (subject) for:

[0224] ;

[0225] in Indicates the first The sample (subjects) and the first The edge weights between samples (subjects).

[0226] Normalized adjacency matrix: Eliminates the impact of node degree differences on convolution, ensuring balanced feature aggregation weights for different subjects.

[0227] ;

[0228] Output the final adjacency matrix ,in For nodes and The normalized edge weights directly reflect the strength of the association between subjects.

[0229] The final graph structure covers all samples, node set for One subject; node feature matrix Adjacency matrix The figure shows an undirected weighted graph.

[0230] 2. Graph Convolutional Networks

[0231] In this embodiment, the graph convolutional network is a graph convolutional network that integrates Neurodynamic-Enhanced and Disentangled representation techniques, named Neurodynamic-Enhanced Disentangled GCN, denoted as NED-GCN.

[0232] The brain region association features are captured through graph convolution, with the following layer structure:

[0233] (1) First-layer graph convolution extracts primary association features: Input Normalized adjacency matrix .

[0234] Convolution calculation: Combining inter-subject similarity and subject characteristics, primary association features are extracted.

[0235] ;

[0236] in These are learnable convolutional weights.

[0237] (2) Second layer graph convolution deepens the associated features: input the first layer convolution features Normalized adjacency matrix .

[0238] Convolutional computation: Further explores deep connections between subjects and compresses feature dimensions to reduce complexity.

[0239] ;

[0240] in These are learnable weights.

[0241] 3. Classifier: Directly outputs the predicted probabilities of each item based on the subject-level convolutional features.

[0242] (1) Graph pooling layer: Aggregates features into global features at the subject level, eliminates redundancy in local features, and retains core correlation information. The formula is as follows: , This represents the graph pooling operation. and The dimensions are consistent, only the features are the values ​​after global aggregation;

[0243] (2) Fully connected layer (feature mapping): Mapping to the classification dimension:

[0244] ;

[0245] in In this method, the learnable weights of the fully connected layer are binary classification weights, with an output dimension of 2. For the bias term, adjust the feature distribution offset. Output the classification feature matrix for the fully connected layer.

[0246] (2) Through Perform probability transformation to obtain the predicted probability:

[0247] ;

[0248] Output prediction probability , [n,0] represents the probability that the sample is a healthy control. [n,1] represents the probability that the sample is an ADHD patient.

[0249] Example 6

[0250] This embodiment, based on embodiments 1-5, further provides an ADHD diagnostic method that integrates neurodynamic enhancement and domain adaptation.

[0251] include:

[0252] S1. Collect data from subjects at multiple sites based on functional magnetic resonance imaging and divide the data into training and validation sets;

[0253] S2. Input the training set data into the ADHD diagnostic system that integrates neurodynamic enhancement and domain adaptation, perform end-to-end training using the total loss function, and evaluate the system's predictive performance using the validation set.

[0254] S3. Apply the qualified system to the diagnosis of the probability of a subject having ADHD.

[0255] The total loss function is:

[0256] ;in, Cross-entropy classification loss measures the difference between the predicted ADHD probability and the true label. Graph structure constraint loss measures the difference between the predicted sample similarity matrix and the true sample similarity matrix. To reduce the association between features and domain labels in the domain adaptation loss, the association with task labels is preserved. and They are respectively and The weighting coefficients.

[0257] ;

[0258] ;in This is the sample similarity matrix predicted by the system. for Norm;

[0259] ;in The first layer of graph convolution features, The balance coefficient is 0.5. Approximated using the InfoNCE estimator.

[0260] Training strategy: The Adam optimizer is used with a learning rate of 0.001, a batch size of 32, and 100-200 iterations. Early stopping is used to prevent overfitting. The evaluation metrics are accuracy, sensitivity, and specificity.

Claims

1. An ADHD diagnostic system that fuses neurodynamic enhancement with domain adaptation, characterized in that, The system includes: Data Input and Preprocessing Module: Receives ROI time-series data from multiple sites based on functional magnetic resonance imaging (fMRI) data from subjects. The sample similarity matrix is ​​obtained through basic feature construction, adaptive affinity fusion, and sample similarity matrix construction. ; Basic feature construction: From Extract the mean and standard deviation of each ROI to obtain the first... The first sample Time series mean of each ROI , and the The first sample Time series standard deviation of ROI ; and Together constitute ,in, For the sample size, Number of brain regions; Adaptive affinity fusion: based on Gaussian kernel similarity is used to construct the feature similarity matrix. , ;based on Pearson correlation coefficients are used to construct functional connectivity matrices. ;Will and Perform fusion to obtain fusion characteristics ; The neural quality-chaotic perturbation module uses a neural quality model to perform basic brain region dynamics simulations on several ROI time series sets, obtaining the simulated data. A nonlinear perturbation signal is generated using a Rössler chaotic system; the nonlinear perturbation signal is then injected... Perform chaotic dynamics fusion to obtain enhanced time series ;from Eight high-discrimination features were extracted to form a dynamic feature matrix. ; The eight high-discrimination features include: the maximum Lyapunov index. Sample entropy Hurst index Low, medium and high frequency energy , and Disturbance variance and local synchronization ; Dynamic characteristic matrix ,in The total number of dynamic features; Domain adaptation module with minimized mutual information: raw image feature matrix of functional magnetic resonance imaging data Based on logistic regression, the model iteratively removes features with the lowest importance until the feature dimension is reached. Stop and output the feature matrix after iteration. ; Multimodal Graph Convolutional Network Module: By constructing a graph structure... and Perform node feature fusion to obtain the node feature matrix. And construct an adjacency matrix It reflects the strength of the association between nodes; it captures the relationship through graph convolution. Brain region association features in the middle, output compressed association features ; By classifying Mapping to the classification dimension yields the classification feature matrix. ,pass right Perform probability distribution transformation to output the probability that the subject is an ADHD patient.

2. The ADHD diagnostic system integrating neurodynamic enhancement and domain adaptation according to claim 1, characterized in that, ROI time-series data from multi-site subjects based on functional magnetic resonance imaging data ,in For the sample size, For time points, Enter the number of brain regions and their corresponding task tags. and domain tags , For healthy controls, For patients with ADHD, domain labels are used to mark collection sites.

3. The ADHD diagnostic system integrating neurodynamic enhancement and domain adaptation according to claim 2, characterized in that, The construction of the sample similarity matrix is ​​as follows: S331, Sample-level affinity graph extension: Calculate the affinity matrix for each sample separately, and obtain... : ； S332, Sample-level statistical feature extraction: Take the mean of all ROI features for each sample to obtain... : ； in For the sample The aggregated mean value of all ROIs For the sample Aggregate standard deviation of all ROIs; S333, Sample Similarity Matrix: The sample similarity matrix is ​​calculated using Gaussian kernel similarity. value range , Indicates sample With sample The similarity.

4. The ADHD diagnostic system integrating neurodynamic enhancement and domain adaptation according to claim 3, characterized in that, In the domain adaptation module that minimizes mutual information, a mutual information definition operation is also performed for features. ,in, Representation of features Feature dimension number, domain label and task tags Define the mutual information of features and domain labels respectively. And mutual information between features and task labels : ； ； in, This represents the total number of all sites participating in the data collection; express Take a specific feature value and domain tags The probability of taking a specific value at a particular site. express Take a specific feature value The probability of that time. Indicates field label Take a specific site value The probability of that time; express Take a specific feature value And task tags The probability of taking a specific probability value. Indicates task tags Take a specific probability value The probability of that time.

5. The ADHD diagnostic system integrating neurodynamic enhancement and domain adaptation according to claim 4, characterized in that, In multimodal graph convolutional networks, node feature fusion specifically involves: Each node represents a subject sample, and the node features are image features. With dynamic characteristics: The integration; First, perform the aggregation operation: for each subject sample... The global features are obtained by averaging the features of all ROIs. Global image features ; Global dynamics ; Then, feature fusion is performed, stitching together the global image features and global dynamic features of the subject samples according to their dimensions to obtain the node feature matrix: ； in For the first The feature vectors of each node have the same dimensions as the selected features in the document. In multimodal graph convolutional networks, the adjacency matrix is ​​constructed. Specifically: Calculate the degree matrix: This calculates the total edge weights for each node, reflecting the global interconnectivity of the nodes. ； ； It is the identity matrix; in, Degree matrix The Middle The sum of the total edge weights of the nodes; Normalized adjacency matrix: Eliminates the impact of node degree differences on convolution, ensuring balanced feature aggregation weights for different subjects. ； Output the final adjacency matrix ,in For nodes and Normalized edge weights.

6. The ADHD diagnostic system integrating neurodynamic enhancement and domain adaptation according to claim 5, characterized in that, Capture through graph convolution Brain region association features in the middle, output compressed association features Specifically: Layer 1 graph convolution extracts primary association features: Input and By combining inter-subject similarity and subject characteristics through convolutional computation, primary association features are extracted. : ； in These are learnable weights; The second layer of graph convolution deepens the associated features: inputting the features from the first layer of convolution. and Further deep connections between participants are explored through convolutional computation, feature dimensions are compressed to reduce complexity, and compressed connection features are output. : ； in These are learnable weights.

7. The ADHD diagnostic system integrating neurodynamic enhancement and domain adaptation according to claim 6, characterized in that, The classifier uses graph pooling layers and fully connected layers to... Mapping to the classification dimension to obtain the classification feature matrix ,pass Output prediction probability , This represents the probability that the sample is a healthy control. This represents the probability that the sample consists of ADHD patients.

8. An ADHD diagnostic method integrating neurodynamic enhancement and domain adaptation, characterized in that, The method employs the ADHD diagnostic system that integrates neurodynamic enhancement and domain adaptation as described in claim 7, comprising: S1. Collect data from subjects at multiple sites based on functional magnetic resonance imaging and divide the data into training and validation sets; S2. Input the training set data into the ADHD diagnostic system that integrates neurodynamic enhancement and domain adaptation, perform end-to-end training using the total loss function, and evaluate the system's predictive performance using the validation set. S3. Apply the qualified system to the diagnosis of the probability that the subject has ADHD.

9. The ADHD diagnostic method integrating neurodynamic enhancement and domain adaptation according to claim 8, characterized in that, The total loss function is: ;in, Cross-entropy classification loss measures the difference between the predicted ADHD probability and the true label. Graph structure constraint loss measures the difference between the predicted sample similarity matrix and the true sample similarity matrix. To reduce the association between features and domain labels, the domain adaptation loss is used. and They are respectively and The weighting coefficients.

Images