A multimodal brain disease diagnostic system based on phenotypic priors and dual-spectral domain synergistic enhancement

Abstract

A multimodal brain disease diagnostic system based on phenotypic priors and dual-spectral-domain synergistic enhancement. This system belongs to the interdisciplinary field of artificial intelligence and brain science. It addresses the technical problem that existing methods struggle to meet practical needs in terms of diagnostic accuracy and model robustness when dealing with highly heterogeneous multicenter clinical data. The system described in this invention preserves the spectral heterogeneity of rs-fMRI signals to enhance feature expression; constructs a more reliable population graph through diagnostic conflict suppression; aligns cross-modal representations using contrastive learning and gating mechanisms; and captures multi-scale population structure by combining spectral-domain graph filtering, thereby significantly improving the accuracy, robustness, and interpretability of brain disease diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of interdisciplinary technology of artificial intelligence and brain science, specifically involving a multimodal brain disease diagnostic system based on phenotypic priors and dual-spectral domain synergistic enhancement. Background Technology

[0002] Resting-state functional magnetic resonance imaging (rs-fMRI), by measuring the correlation of spontaneous neural activity between brain regions, can reflect the intrinsic functional connectivity patterns of the brain and has been widely used in the auxiliary diagnosis of brain disorders such as autism spectrum disorder (ASD) and attention deficit hyperactivity disorder (ADHD). Existing graph-based brain disease prediction methods typically construct a graph of the subject or brain regions and use graph neural networks for feature aggregation and classification. However, existing methods generally suffer from the following drawbacks: 1. Loss of frequency information: Most existing research methods use full-band / single wideband to calculate functional connectivity, ignoring the significant frequency dependence of rs-fMRI signals. Different frequency subbands correspond to different neural oscillation scales and functional tissue patterns, and disease-related abnormalities often only appear in specific frequency bands. Single-band analysis leads to over-smoothing of fine-grained pathological features, failing to capture disease-specific patterns.

[0003] 2. Population graph noise: When constructing a population graph among subjects, connections are usually established based on age, sex, or phenotypic similarity at the collection sites. However, subjects with similar phenotypes may have completely different diagnostic labels. If such "diagnostic conflict" relationships are treated as positive edges for information propagation, it will introduce erroneous information aggregation and severely weaken the discriminative and generalization abilities of the graph structure.

[0004] 3. Cross-modal inconsistency: Image features and phenotypic features differ greatly in statistical distribution and semantic space. Existing simple feature splicing or fixed weight fusion methods cannot guarantee the alignment of the two modalities in the embedding space, resulting in semantic shift between modalities and affecting the robustness of the model.

[0005] 4. Lack of multi-scale structure: Traditional graph convolution only performs neighborhood aggregation in the spatial domain, which is prone to oversmoothing and cannot simultaneously capture the global population structure and local heterogeneous subpopulations in the population graph, resulting in insufficient ability to model complex disease distributions.

[0006] The aforementioned shortcomings mean that existing methods struggle to meet practical needs in terms of diagnostic accuracy and model robustness when faced with highly heterogeneous multicenter clinical data. Summary of the Invention

[0007] To address the core shortcomings of existing technologies, such as frequency information loss, population graph diagnostic conflicts, cross-modal inconsistencies, and missing multi-scale structures, this invention provides a multimodal brain disease diagnostic system based on phenotypic priors and dual-spectral domain synergistic enhancement. The system aims to achieve the following objectives: preserve the spectral heterogeneity of rs-fMRI signals to enhance feature expression; construct a more reliable population graph through diagnostic conflict suppression; align cross-modal representations using contrastive learning and gating mechanisms; and capture multi-scale population structures by combining spectral domain graph filtering, thereby significantly improving the accuracy, robustness, and interpretability of brain disease diagnosis.

[0008] The system specifically includes: Multi-band functional connectivity feature extraction and fusion unit: Receives rs-fMRI brain region time series and phenotypic data of subjects; performs multi-band decomposition on rs-fMRI brain region time series to obtain time series under each frequency band, constructs functional connectivity similarity matrix for each frequency band, introduces learnable frequency band weights to adaptively fuse multi-band similarity, and obtains spectral functional similarity matrix; Relationship Consistency Population Graph Construction Unit: Similarity discrimination is performed on phenotypic data, support relationship matrix and conflict relationship matrix are constructed, then relationship consistency graph is constructed and fused to obtain phenotypic relationship graph, and the spectral functional similarity matrix and phenotypic relationship graph are combined to obtain the final population graph used for image view coding; Cross-modal graph representation learning and dual-spectral-domain collaborative enhancement of the classification unit: A shared graph encoder is constructed, and the image view and phenotypic view are input. The graph encoder with shared parameters obtains the subject representations encoded by the image view and the subject representations encoded by the phenotypic view. The two are adaptively gated and fused to obtain the fused subject node representation. Multi-branch spectral-domain enhancement is performed based on the subject node representation and the population graph to obtain the enhanced subject representation. The enhanced subject representation is input into the classifier to obtain the brain disease prediction result.

[0009] Furthermore, the rs-fMRI brain region time series were decomposed into multiple frequency bands to obtain the time series for each frequency band as follows: rs-fMRI brain region time series of each subject The bandpass filtering for the frequency band is specifically as follows: for the Brain region time series of individual subjects Through the first The bandpass filter yields the first... Brain region time series of individual subjects : ;in, Indicates the first Filtering operation for each frequency sub-band.

[0010] Furthermore, the functional connectivity similarity matrix for each frequency band is constructed as follows: For the time series under each frequency band The frequency band-specific function connectivity matrix is ​​calculated as follows: ; in, Represents the inverse hyperbolic tangent function. Indicates the first Brain region time series under each frequency band Calculate the Pearson correlation coefficient matrix; Further take The upper triangular elements are quantized to obtain the first... Functional connectivity features under each frequency band Functional connectivity features across different frequency bands collectively constitute the subject's spectral perception image representation: ;in, Indicates feature splicing; For the Each frequency band, subjects and The functional connectivity similarity matrix between them is defined as follows: ; in, For the first Scale parameters of inter-sample distance in each frequency band The operation represented is an exponential operation with the natural constant e as the base. The operation represented is the squaring of the Euclidean norm; Representation matrix The Middle Okay, number Column elements, Indicates the first Functional connectivity similarity matrix for each frequency band.

[0011] Furthermore, the spectral functional similarity matrix for: , ; in, For the first Learnable parameters for each frequency band, Indicates the first Normalized weights for each frequency band.

[0012] Furthermore, similarity is determined from the phenotypic data, and support and conflict relation matrices are constructed as follows: For any phenotypic variable The similarity determination method is as follows: for discrete variables, if two subjects take the same value, they are considered similar in that phenotype; for continuous variables, similarity is determined based on whether the difference between the two values ​​is less than a given threshold, and the phenotypic similarity is denoted as... ; Construct supporting relation matrix elements Elements of the conflict relationship matrix : ; ; This leads to the support relation matrix. Conflict Relationship Matrix Its matrix elements are respectively derived from the aforementioned and definition.

[0013] in, Indicates the subject Diagnostic label, Indicates the subject Diagnostic labels.

[0014] Furthermore, the relational consistency graph is defined as: ; in, This is the inhibition coefficient of conflict relations. Represents a nonlinear normalization function. Represents phenotypic variables Next subject and Unlabeled relationships between them; The phenotypic relationship diagram is as follows: ; in, Indicates the first Normalized weights of phenotypic variables, phenotypic variables The corresponding learnable parameters, This represents the number of phenotypic variables.

[0015] Furthermore, a shared graph encoder is constructed, using multi-band functional connectivity features as image view input. The phenotypic variables, after feature mapping, are used as inputs for the phenotypic view. Both are projected onto a unified dimensional space through linear mapping: ; ; express Initial feature representation in the encoder input layer for Initial feature representation at the encoder input layer; and A learnable projection matrix is ​​then used; subsequently, the two views are input into a graph encoder with shared parameters for message passing. ; ; in, Indicates a shared graph encoder. A population graph used for image view encoding. and These represent the subject representations obtained from image view and phenotypic view encodings, respectively.

[0016] Furthermore, the fused subject nodes are represented as follows: ; in, This represents the fused subject node representation. and These are the gating weights.

[0017] Furthermore, multi-branch spectral domain enhancement is performed based on the subject node representation and population graph to obtain the enhanced subject representation as follows: Calculate the population graph Normalized graph Laplace matrix A spectral domain graph filter is constructed based on the Chebyshev polynomial approximation: ; in, Indicates the first The subject representation from the output of each spectral filter branch, This represents the scaled graph Laplacian matrix. for Chebyshev polynomial of order 1 For the first Learnable parameters in each spectral filtering branch are used to concatenate and fuse the outputs of multiple branches to obtain the enhanced subject representation. : ; in, Represents a nonlinear fusion function. This represents the number of branches in the spectral filter.

[0018] Furthermore, by inputting the enhanced subject representations into the classifier, the brain disease prediction results are obtained as follows: ; This represents the disease prediction results presented in probabilistic form, that is, the model's predicted probability distribution for each category. This represents a multilayer perceptron classifier. This represents the normalized exponential function.

[0019] The beneficial effects of the system described in this invention are as follows: 1. Significantly improved diagnostic accuracy: On public datasets such as ABIDE, the accuracy and AUC of this system outperform many existing graph learning methods, especially showing stable performance on multi-center heterogeneous data.

[0020] 2. More granular feature representation: Multi-band functional connectivity preserves disease-related local abnormality patterns that are masked by full-band smoothing, making the description of brain functional tissue more accurate.

[0021] 3. More reliable population structure: Conflict relationship suppression effectively avoids erroneous aggregation among subjects with inconsistent diagnoses, making the graph topology more consistent with the actual disease distribution and reducing noise propagation.

[0022] 4. Cross-modal representation alignment: Contrastive learning and gating fusion ensure that image and phenotypic embeddings are highly consistent in the same space, improving the credibility and robustness of fused features.

[0023] 5. Multi-scale topological perception: The spectral domain enhancement module simultaneously captures global population consistency and local subgroup boundaries, enabling it to adapt to the complex subtyping characteristics of brain diseases.

[0024] 6. Enhanced model robustness: The gating mechanism can dynamically reduce the contribution of low-quality modes, and together with graph regularization constraints, it can effectively resist the interference of missing data, noise and site differences. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating the workflow of the system described in this embodiment of the invention. Detailed Implementation

[0026] 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.

[0027] Example 1 This embodiment provides a multimodal brain disease diagnostic system based on phenotypic prior and dual-spectral domain synergistic enhancement. The workflow diagram of the system is as follows: Figure 1 As shown, the brain disease diagnostic system in this embodiment is based on multimodal imaging and phenotypic data, and is implemented through three core stages: multi-band functional connectivity representation, relational consistency population graph construction, and cross-modal graph representation learning and dual-spectral domain collaborative enhancement classification. The specific unit structure is as follows: Multi-band functional connection feature extraction and fusion unit: Given contains The dataset of the nth subject, denoted as the nth subject... The rs-fMRI time series of each subject is as follows: ,in Indicates the number of brain regions. Indicates the number of time points; its corresponding phenotypic information is: Diagnostic label is The rs-fMRI brain region time series of each subject were analyzed. Bandpass filtering was applied to each frequency band (slow-5, slow-4, and slow-3 in this embodiment). Pearson correlation coefficients were calculated for brain regions within each frequency band, and Fisher's z-transform was performed to obtain a frequency-specific functional connectivity similarity matrix. The upper triangular elements were extracted and quantized to obtain the frequency-specific functional connectivity features. Subsequently, a similarity matrix between subjects was calculated based on the features of each frequency band, and learnable frequency band weights were introduced for adaptive weighted fusion to obtain a spectrum-aware fusion similarity matrix. .

[0028] First, the brain region time series of each subject was decomposed into multiple frequency bands. For the first... Brain region time series of individual subjects Through the first Each bandpass filter yields the signal representation for the corresponding frequency band: ; in, Indicates the first Filtering operations for each frequency sub-band. This embodiment uses multiple low-frequency sub-bands to divide the rs-fMRI signal to characterize functional connectivity patterns within different frequency ranges.

[0029] Subsequently, for the time series under each frequency band Calculate the Pearson correlation coefficient matrix between brain regions and perform Fisher's algorithm. Transformation yields the frequency band-specific function connectivity similarity matrix: ; in, This represents the inverse hyperbolic tangent function, also known as the Fisher-Z transform, which is used to standardize the correlation coefficient matrix. Indicates the first Brain region time series under each frequency band Calculate the Pearson correlation coefficient matrix; Further, take its upper triangular elements and quantize them to obtain the first... Functional connectivity features under each frequency band Functional connectivity features across different frequency bands collectively constitute the subject's spectral perception image representation: ;in, This indicates feature splicing.

[0030] To leverage functional similarity across different frequency bands at the group level, this embodiment further calculates a similarity matrix among subjects based on the functional connectivity features of each frequency band. For the first... Each frequency band, subjects and The similarity between them is defined as: ; in, For the first Scale parameters of inter-sample distance in each frequency band The operation represented is an exponential operation with the natural constant e as the base. The operation represented is the squaring of the Euclidean norm. Indicates the first The subject and the first The first subject was on the The squared Euclidean distance between the functional connectivity features of each frequency band is used to measure the degree of difference in the frequency band features of two subjects.

[0031] Considering that different frequency bands may contribute differently to disease prediction, this embodiment introduces learnable frequency band weights to adaptively fuse multi-frequency band similarity to obtain a spectral functional similarity matrix. : , ; in, Indicates the first Functional connectivity similarity matrix for each frequency band; For the first Learnable parameters for each frequency band, Indicates the first Normalized weights for each frequency band. Through this design, the model can automatically emphasize more disease-discriminating frequency band information according to task requirements, thereby improving the fine-grained expressive power of functional connectivity representation; and Indicates different levels, : No. Functional connectivity similarity matrix for each frequency band; Representation matrix The Middle line, number The element of the column, that is, the first The first subject and the first The first subject was on the Functional connectivity similarity across frequency bands; The two terms represent the overall and local relationships of similarity within the same frequency band.

[0032] Relationship Consistency Population Graph Construction Unit: Population graph learning plays a crucial role in predicting brain diseases by modeling the relationship structure between subjects. Traditional methods typically establish subject connections based on phenotypic similarity such as age, gender, and data collection site. However, phenotypic similarity does not necessarily imply diagnostic consistency. Directly treating all phenotypic similarity relationships as positive edges for information propagation may lead to erroneous aggregation of subjects with conflicting diagnoses. To address this, this embodiment proposes a relationship-consistent population graph construction strategy, further introducing diagnostic consistency constraints on top of phenotypic similarity.

[0033] For any phenotypic variable First, we determine whether the subjects are similar based on their phenotypic attributes.

[0034] The specific determination method is as follows: For discrete variables, such as gender or data collection site, two subjects are considered similar in phenotype if they have the same value. For continuous variables, such as age, similarity is determined by whether the difference between the two values ​​is less than a given threshold. This phenotypic similarity is denoted as […]. .

[0035] For discrete phenotypes (such as gender, data collection site), then: Subjects and In phenotypic variables The values ​​above are the same; Subjects and In phenotypic variables The values ​​taken above are different.

[0036] For continuous phenotypes (such as age), a given threshold is given. ,but: Subjects and In phenotypic variables The absolute value of the difference between the values ​​on is less than ; Subjects and In phenotypic variables The absolute value of the difference between the values ​​is greater than or equal to In this embodiment, the age threshold is set to... .

[0037] For missing values, then: .

[0038] In the training set, if two subjects have similar phenotypes and identical diagnostic labels, their relationship is defined as a support relationship; if they have similar phenotypes but different diagnostic labels, their relationship is defined as a conflict relationship. Correspondingly, a support relationship matrix is ​​constructed. Conflict Relationship Matrix : ; ; in, Indicates the subject Diagnostic label, Indicates the subject Diagnostic labels; For subjects in the validation and test sets, since their labels cannot participate in graph construction, this embodiment only establishes an unlabeled relation matrix based on phenotypic similarity. To avoid label leakage.

[0039] Represents phenotypic variables Next subject and Unlabeled relationships between subjects, used for subject pairs involving validation or test sets, are constructed without diagnostic labels.

[0040] ={1, Subjects or It belongs to the validation set or the test set, and both are phenotypic variables. The following conditions must be met for a tagless connection to be established; otherwise... =0} Conflict relationships are not directly input into the graph neural network as negative edges. Since negative edges can disrupt the stability of the graph Laplacian and make the message propagation direction difficult to interpret, this embodiment models conflict relationships as propagation strength suppression terms during the topology construction phase. Specifically, for phenotypic variables... Its relationship consistency graph is defined as: ; in, This is the inhibition coefficient of conflict relations. This represents a nonlinear normalization function used to map relation weights to a stable propagation range. Instead of directly removing conflicting edges, this design reduces their propagation strength during topology construction, thereby mitigating the impact of diagnosing inconsistencies in relations on information aggregation.

[0041] and Representing phenotypic variables respectively The corresponding support and conflict matrices have their matrix elements derived from the aforementioned... and definition; Represents phenotypic variables The corresponding unlabeled relation matrix.

[0042] Different phenotypic variables may contribute differently to disease prediction; therefore, this embodiment further introduces phenotypic weights to fuse multiple relationship consistency graphs: ; in, Indicates the first Normalized weights of phenotypic variables, phenotypic variables The corresponding learnable parameters, This represents the number of phenotypic variables.

[0043] Finally, this embodiment will use the spectral functional similarity matrix. Relationship with phenotype diagram By combining these elements, we obtain the final population graph used for image view encoding: ; This operation subjectes the final population graph to two constraints simultaneously: phenotypic relationships determine which subjects have potential group connections, while spectral functional similarity further modulates the strength of these connections. The resulting relation-consistent population graph can suppress the propagation of errors caused by diagnostic conflict relationships while preserving the population topology.

[0044] Cross-modal graphical representation learning and dual-spectral domain collaborative enhancement of classification units: After obtaining the relational consistency population graph and phenotypic prior graph, this unit completes end-to-end learning from multimodal features to disease prediction results through four sub-stages: shared graph encoding, cross-modal contrastive alignment and gating fusion, spectral domain enhancement, and classification.

[0045] (1) Construct a shared graph encoder First, the model uses multi-band functional connectivity features as image view input, and uses phenotypic variables, after feature mapping, as phenotypic view input. Let the image feature matrix be... The phenotypic feature matrix is Both are projected onto a unified dimensional space through linear mapping: ; ; Representation of the characteristic matrix Initial feature representation in the encoder input layer Phenotypic feature matrix Initial feature representation at the encoder input layer; and This is a learnable projection matrix. Subsequently, the two views are input into a graph encoder with shared parameters for message passing (using a multi-layered TransformerConv): ; ; in, Indicates a shared graph encoder. This is a population graph showing the consistency of spectrum-sensing relationships. This is a phenotypic relationship graph. The shared encoder ensures that image embeddings and phenotypic embeddings are mapped to the same semantic space, fundamentally reducing intermodal shifts; and These represent the subject representations obtained from image view and phenotypic view encodings, respectively.

[0046] (2) Cross-modal contrastive learning and gating fusion To force alignment of the same subject's representation across different modalities, a cross-modal contrast loss is introduced. For the same subject, their image embeddings... With phenotypic embedding Cross-modal embeddings between different subjects are considered positive sample pairs; while cross-modal embeddings between different subjects are considered negative sample pairs. Contrast loss is defined as: ; in, Represents the similarity function. The temperature coefficient is used. This constraint makes cross-view representations of the same subject attract each other, while representations of different subjects repel each other, thereby enhancing the discriminative structure of the representation space. Finally, the model adaptively integrates the image view and phenotypic view through a gating fusion mechanism to obtain the fused subject node representation. : ; in, and These represent the gating weights for the image view and the phenotypic view, respectively. Compared to direct stitching, this fusion method can dynamically adjust the contributions of the two types of information based on the modal reliability of different subjects, avoiding excessive interference from low-quality modalities in the fusion result.

[0047] (3) Dual-spectral domain synergistic enhancement While relational consistency population graphs can mitigate the propagation of erroneous relationships, the group relationship patterns in complex brain diseases often exhibit multi-scale characteristics. Local neighborhood structures may reflect fine-grained disease differences among similar subjects, while larger-scale graph structures may correspond to global distribution patterns among different disease subgroups. Therefore, this embodiment introduces a spectral domain graph enhancement module onto the fused subject representation to capture multi-scale topological responses in the population graph.

[0048] Given the fused subject representation and population map First, calculate the population graph. Normalized graph Laplace matrix Subsequently, a spectral domain graph filter was constructed based on the Chebyshev polynomial approximation: ; in, Indicates the first The subject representation output of each spectral filter branch, that is, different Spectral filters corresponding to different scales or frequency bands are used to capture graphical responses of different ranges; This represents the scaled graph Laplacian matrix. for Chebyshev polynomial of order 1 For the first Learnable parameters in each spectral filtering branch. Different spectral filtering branches are used to capture graphical structure responses at different scales. Subsequently, the outputs of multiple branches are concatenated and fused to obtain the enhanced subject representation: ; in, Represents a nonlinear fusion function. This represents the number of spectral filter branches. This module enables the model to extract multi-scale structures in the population graph domain, further enhancing its ability to model complex population relationships.

[0049] The multi-band functional connectivity (temporal spectrum) in the multi-band functional connectivity feature extraction and fusion unit and the Chebyshev spectral domain enhancement (spatial map) in this unit operate on two different dimensions: the former enhances the functional connectivity representation from the temporal frequency dimension of the rs-fMRI signal, while the latter enhances the node embedding representation from the topological spectrum dimension of the population graph. Together, they constitute the core mechanism of "dual-spectral domain perception".

[0050] (4) Classifier and joint optimization Ultimately, the enhanced subjects indicated that inputting the classifier yielded disease prediction results: ; This represents the disease prediction results presented in probabilistic form, that is, the model's predicted probability distribution for each category. It means: representing the subjects after spectral domain enhancement. Input the data into a multilayer perceptron classifier to obtain a classification score, and then... The function is converted into class probabilities.

[0051] in, This represents a multilayer perceptron classifier used to map enhanced subject representations to classification scores for each disease category; This represents the normalized exponential function, used to convert classification scores into predicted probabilities for each category. The model's classification loss uses cross-entropy loss. ; in, This represents the set of subjects in the training set. Indicates the first The prediction results of each subject.

[0052] To further stabilize representation learning on the population graph, this embodiment introduces a graph regularization term on the relation-consistent population graph: ; in, Used to constrain the consistency of subject representations within a reliable neighborhood. To prevent excessive aggregation of supernodes in the graph, a degree penalty term is added. and where is the weighting coefficient. This smoothing constraint does not apply to the original phenotypic similarity graph, but rather to the relation consistency graph after conflict relation suppression. Therefore, its purpose is not to unconditionally enhance neighborhood smoothness, but to further improve the representation stability of locally reliable relations after reducing diagnostic conflict propagation.

[0053] Therefore, the overall optimization objective is: ; in, The classification and discrimination ability of the constraint model. Constrain cross-modal consistency between image views and phenotypic views. This is used to maintain the representation stability and topological constraints of population graphs with consistent relationships; These are the weighting coefficients.

[0054] By jointly optimizing the above objectives, it is possible to learn subject representations that combine spectral sensitivity, relational consistency, and cross-modal robustness.

[0055] Example 2 This embodiment is a further explanation of Embodiment 1, which helps to better understand the technical solution in Embodiment 1.

[0056] Working principle explanation: The core working principle of Example 1 is to guide population graph construction with phenotypic priors and achieve multimodal brain disease diagnosis through synergistic enhancement of the temporal and spatial spectral domains: 1. Temporal spectrum sensing: Multi-band decomposition of rs-fMRI time series preserves the specific functional connectivity patterns of different frequency bands, solving the problem of insufficient single-scale feature expression.

[0057] 2. Phenotypic Prior Guidance: Diagnostic consistency is introduced as an explicit constraint into population graph construction. By distinguishing between supportive and conflicting relationships and imposing inhibition, the risk of erroneous topology propagation is fundamentally reduced.

[0058] 3. Cross-modal contrast alignment: In the shared graph encoder, contrast loss is used to bring the cross-modal representations of the same subject closer together, so as to achieve adaptive fusion of images and phenotypes in a unified semantic space.

[0059] 4. Spatial Spectral Domain Enhancement: The joint embedding is refined in the multi-scale spectral domain using a graph wavelet Chebyshev filter, while preserving the global smoothness at the population level and the discrimination boundaries of local subpopulations.

[0060] 5. Gated quality perception: Modal gating adaptively adjusts the dependence of different samples on image and phenotypic information, further improving diagnostic robustness.

[0061] Key technical points explained: 1. Multi-band functional connection and adaptive fusion: For the first time, RS... fMRI is decomposed into multiple time-frequency subbands, and multi-band subject similarity is adaptively fused through learnable weights to enhance the ability to express disease-related spectral heterogeneity.

[0062] 2. Population graph construction mechanism for diagnosing conflict perception: It innovatively distinguishes phenotypic relationships into support relationships and conflict relationships, and introduces a conflict suppression coefficient to reduce the propagation intensity of diagnostic inconsistent connections during the topology construction stage, thus purifying the population graph from the source.

[0063] 3. Cross-modal contrastive gating fusion with shared graph encoding: The image and phenotypic representation are forcibly aligned by using a shared graph Transformer encoder and cross-modal contrastive loss, and a gating mechanism is used to achieve sample-level modal dynamic reconstruction, avoiding semantic shifts in simple fusion.

[0064] 4. Dual-spectral domain collaborative enhancement architecture: Simultaneously model the temporal spectrum (multi-band) and spatial spectrum (Chebyshev wavelet graph), use low-order spectral filtering to maintain the commonality of the population, and use high-order spectral response to capture local anomalies, so as to realize multi-scale characterization of complex population relationships.

[0065] 5. End-to-end multi-objective joint training strategy: It integrates classification loss, cross-modal contrastive loss and graph regularization loss, while constraining discriminative power, cross-modal consistency and population topological stability, so that the model can reach the optimal state without step-by-step processing.

[0066] System implementation details: This paper implements a platform application for the multimodal brain disease diagnostic system based on phenotypic priors and dual-spectral domain synergistic enhancement using PyTorch and PyTorch Geometric. End-to-end training is performed using the Adam optimizer. All experiments were conducted on the same hardware environment to ensure fairness in comparisons between different methods. During system model training, the initial learning rate was set to... The weight decay coefficient is set to The maximum number of training epochs was set to 500. To mitigate overfitting, Dropout was employed in both the graph encoder and classifier, with a dropout rate of 0.3. Simultaneously, during training, edge connections in the population graph were randomly dropped, also with a dropout rate of 0.3. An early stopping strategy was used: training was stopped when the validation set loss no longer decreased over 100 consecutive epochs, and the model with the best performance on the validation set was saved for testing.

[0067] In terms of feature processing, feature selection is first performed on high-dimensional functional connectivity features based on the training set, and the dimension of image features is uniformly reduced to 500 dimensions to reduce the impact of redundant connections on model training. For phenotypic information, non-image variables such as age, gender, and acquisition site are encoded as numerical features and mapped to a representation space consistent with image features for subsequent cross-modal graph encoding. It should be noted that all information involving training labels is only used for constructing the relation consistency graph in the training set; the diagnostic labels of the validation and test set samples are not used during graph construction to avoid label leakage.

[0068] Regarding the system's network architecture, cross-modal graph representation learning employs a shared-parameter graph Transformer structure to encode image views and phenotypic views separately. The graph encoder's hidden dimension is set to 128, and the output dimension to 16. Multi-branch spectral domain augmentation is implemented using Chebyshev polynomial-based graph spectral filtering, with the Chebyshev polynomial order set to 4, and four parallel spectral filtering branches used to capture graph structural responses at different scales. For multi-band functional connectivity modeling, three frequency sub-bands are used, and functional similarity across different frequency bands is adaptively fused using learnable weights.

[0069] Specific applications: The system in Example 1 was validated on the ABIDE dataset, which contains multimodal neuroimaging data from 403 patients with autism spectrum disorder (ASD) and 468 healthy controls (HC). The specific implementation parameters and results are shown in Table 1 below: Table 1:

[0070] Application effect: 1. Diagnostic performance: Example 1 achieved an average accuracy of 89.0% and an AUC of 93.78% with 10-fold cross-validation on the ABIDE dataset, which is better than the traditional GCN population graph method and single-band method.

[0071] 2. Conflict suppression effect: When conflict suppression is canceled ( The accuracy decreased by about 2%, which verified the key role of relation consistency constraints in the quality of population graphs.

[0072] 3. Multi-band gain: When only full-band functionality is used (multi-band decomposition is disabled), the AUC decreases by approximately 5.3%, confirming that spectrum-sensing modeling can effectively improve disease discrimination.

[0073] 4. Spectral domain enhancement contribution: After removing the Chebyshev enhancement branch (using only single-scale graph convolution), the performance variance of the model on multi-site subsets increased, indicating that multi-scale filtering in the graph domain enhanced the stability under heterogeneous data.

[0074] 5. Cross-modal alignment validation: After adding contrast loss, the mutual information between image and phenotype embeddings is significantly improved, and the classification boundary is clearer.

[0075] Comparison of the effects of different parameter values: 1. Conflict suppression coefficient :when As the value increases from 0 to 1.0, the model accuracy gradually rises. However, after exceeding 1.5, performance declines due to over-penalizing similarity relationships. The optimal range is 0.8. 1.2.

[0076] 2. Number of frequency bands: The performance is slightly lower when using 2 frequency bands than when using 3 frequency bands. Using 4 or more frequency bands does not provide additional gain but increases the computational burden. 3 frequency bands is the best balance point.

[0077] 3. Compare the temperature loss : The positive and negative sample pairs have the best discriminative power. When the value is greater than 0.1, the contrast constraint is too weak. Training is unstable when the value is less than 0.05.

[0078] 4. Edge dropout rate: The test performance is best when the edge dropout rate is 0.3, training is too slow when it is 0.5, and there is slight overfitting when it is 0.

Claims

1. A multimodal brain disease diagnostic system based on phenotypic prior and dual-spectral domain synergistic enhancement, characterized in that, The system includes: Multi-band functional connectivity feature extraction and fusion unit: Receives rs-fMRI brain region time series and phenotypic data of subjects; performs multi-band decomposition on rs-fMRI brain region time series to obtain time series under each frequency band, constructs functional connectivity similarity matrix for each frequency band, introduces learnable frequency band weights to adaptively fuse multi-band similarity, and obtains spectral functional similarity matrix; Relationship Consistency Population Graph Construction Unit: Similarity discrimination is performed on phenotypic data, support relationship matrix and conflict relationship matrix are constructed, then relationship consistency graph is constructed and fused to obtain phenotypic relationship graph, and the spectral functional similarity matrix and phenotypic relationship graph are combined to obtain the final population graph used for image view coding; Cross-modal graph representation learning and dual-spectral-domain collaborative enhancement of the classification unit: A shared graph encoder is constructed, and the image view and phenotypic view are input. The graph encoder with shared parameters obtains the subject representations encoded by the image view and the subject representations encoded by the phenotypic view. The two are adaptively gated and fused to obtain the fused subject node representation. Multi-branch spectral-domain enhancement is performed based on the subject node representation and the population graph to obtain the enhanced subject representation. The enhanced subject representation is input into the classifier to obtain the brain disease prediction result.

2. The multimodal brain disease diagnostic system based on phenotypic prior and dual-spectral domain synergistic enhancement according to claim 1, characterized in that, Multi-band decomposition was performed on the rs-fMRI brain region time series to obtain the time series for each band as follows: rs-fMRI brain region time series of each subject The bandpass filtering for the frequency band is specifically as follows: for the Brain region time series of individual subjects Through the first The bandpass filter yields the first... Brain region time series of individual subjects : ;in, Indicates the first Filtering operation for each frequency sub-band.

3. The multimodal brain disease diagnostic system based on phenotypic prior and dual-spectral domain synergistic enhancement according to claim 2, characterized in that, The functional connectivity similarity matrix for each frequency band is constructed as follows: For the time series under each frequency band The frequency band-specific function connectivity matrix is ​​calculated as follows: ; in, Represents the inverse hyperbolic tangent function. Indicates the first Brain region time series under each frequency band Calculate the Pearson correlation coefficient matrix; Further take The upper triangular elements are quantized to obtain the first... Functional connectivity features under each frequency band Functional connectivity features across different frequency bands collectively constitute the subject's spectral perception image representation: ;in, Indicates feature splicing; For the Each frequency band, subjects and The functional connectivity similarity matrix between them is defined as follows: ； in, For the first Scale parameters of inter-sample distance in each frequency band The operation represented is an exponential operation with the natural constant e as the base. The operation represented is the squaring of the Euclidean norm; Representation matrix The Middle Okay, number Column elements, Indicates the first Functional connectivity similarity matrix for each frequency band.

4. The multimodal brain disease diagnostic system based on phenotypic prior and dual-spectral domain synergistic enhancement according to claim 3, characterized in that, Spectral Functional Similarity Matrix for: , ; in, For the first Learnable parameters for each frequency band, Indicates the first Normalized weights for each frequency band.

5. The multimodal brain disease diagnostic system based on phenotypic prior and dual-spectral domain synergistic enhancement according to claim 4, characterized in that, Similarity is determined for the phenotypic data, and the support relation matrix and conflict relation matrix are constructed as follows: For any phenotypic variable The similarity judgment method is as follows: for discrete variables, if two subjects take the same value, they are considered to be similar in that phenotype. For continuous variables, similarity is determined by whether the difference between the two is less than a given threshold, and this phenotypic similarity is denoted as... ; Construct supporting relation matrix elements Elements of the conflict relationship matrix : ； ； This leads to the support relation matrix. Conflict Relationship Matrix Its matrix elements are respectively derived from the aforementioned and definition; in, Indicates the subject Diagnostic label, Indicates the subject Diagnostic labels.

6. The multimodal brain disease diagnostic system based on phenotypic prior and dual-spectral domain synergistic enhancement according to claim 5, characterized in that, A relational consistency graph is defined as: ; in, This is the inhibition coefficient of conflict relations. Represents a nonlinear normalization function. Represents phenotypic variables Next subject and Unlabeled relationships between them; The phenotypic relationship diagram is as follows: ; in, Indicates the first Normalized weights of phenotypic variables, phenotypic variables The corresponding learnable parameters, This represents the number of phenotypic variables.

7. The multimodal brain disease diagnostic system based on phenotypic prior and dual-spectral domain synergistic enhancement according to claim 6, characterized in that, Construct a shared graph encoder that uses multi-band functional connectivity features as image view input. The phenotypic variables, after feature mapping, are used as inputs for the phenotypic view. Both are projected onto a unified dimensional space through linear mapping: ； ； express Initial feature representation in the encoder input layer for Initial feature representation at the encoder input layer; and A learnable projection matrix is ​​then used; subsequently, the two views are input into a graph encoder with shared parameters for message passing. ; ; in, Indicates a shared graph encoder. A population graph used for image view encoding. and These represent the subject representations obtained from image view and phenotypic view encodings, respectively.

8. The multimodal brain disease diagnostic system based on phenotypic prior and dual-spectral domain synergistic enhancement according to claim 7, characterized in that, The fused subject nodes are represented as follows: ; in, This represents the fused subject node representation. and These are the gating weights.

9. The multimodal brain disease diagnostic system based on phenotypic prior and dual-spectral domain synergistic enhancement according to claim 8, characterized in that, Multi-branch spectral domain augmentation is performed based on the subject node representation and population graph to obtain the augmented subject representation as follows: Calculate the population graph Normalized graph Laplace matrix A spectral domain graph filter is constructed based on the Chebyshev polynomial approximation: ； in, Indicates the first The subject representation from the output of each spectral filter branch, This represents the scaled graph Laplacian matrix. for Chebyshev polynomial of order 1 For the first Learnable parameters in each spectral filtering branch are used to concatenate and fuse the outputs of multiple branches to obtain the enhanced subject representation. : ； in, Represents a nonlinear fusion function. This represents the number of branches in the spectral filter.

10. The multimodal brain disease diagnostic system based on phenotypic prior and dual-spectral domain synergistic enhancement according to claim 9, characterized in that, The enhanced subject representations were input into the classifier to obtain the brain disease prediction results: ; This represents the disease prediction results presented in probabilistic form, that is, the model's predicted probability distribution for each category. This represents a multilayer perceptron classifier. This represents the normalized exponential function.

Images