A cognitive network key node spatial positioning method based on PET-MR

By employing cross-modal feature alignment and fusion, dynamic cognitive network construction, and spatial guided field localization, the problem of insufficient accuracy in dynamic network modeling and localization in PET-MR images was solved, enabling precise localization and interpretable assessment of key nodes in individualized brain networks.

CN122866321BActive Publication Date: 2026-11-10NANJING FIRST HOSPITAL
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611344986.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-09-01
Publication Date
2026-11-10
Estimated Expiration
2046-09-01

AI Technical Summary

Technical Problem

Existing methods for constructing brain networks based on PET-MR images are insufficient in characterizing dynamic evolutionary features, lack cross-modal feature interactions, have inadequate localization accuracy, and poor interpretability, thus affecting brain science research and analysis.

Method used

By constructing a dynamic cognitive network through cross-modal feature alignment and fusion guided by semantic prototypes, and combining neural modulation and hierarchical community detection, node coordinates are refined using spatial guidance fields and differentiable anatomical constraints, and multi-granularity confidence assessment is performed.

Benefits of technology

It achieves deep fusion of PET-MR multimodal data, accurately locates individualized key nodes, improves the accuracy, stability and interpretability of positioning, and provides multi-dimensional credibility assessment.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122866321B_ABST
    Figure CN122866321B_ABST
Patent Text Reader

Abstract

The application discloses a PET-MR-based cognitive network key node spatial positioning method and relates to the technical field of medical image processing. The method comprises the following steps: acquiring multi-modal images and performing pretreatment; guiding cross-modal feature alignment and fusion by using semantic prototypes to generate high-resolution metabolic maps and edge space confidence maps; constructing a dynamic cognitive network through a global neural modulator; combining hierarchical community detection and space-guided field-driven coordinate offset; constructing a graph neural network to predict spatial residuals; refining coordinates through a differentiable anatomical constraint solver; and finally performing multi-granularity confidence evaluation and error feedback. The application realizes multi-modal deep fusion and individualized accurate positioning, and improves positioning accuracy, stability and result interpretability.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method for spatial localization of key nodes in a cognitive network based on PET-MR. Background Technology

[0002] Integrated PET-MR (Positron Emission Tomography-Magnetic Resonance) equipment can simultaneously acquire PET metabolic images and MR structural images of the brain, providing a multimodal data foundation for brain cognitive network research. Constructing brain networks and locating key nodes based on PET-MR images is of great value for revealing the neural mechanisms of cognitive function and conducting imaging research analysis related to neurodegenerative diseases.

[0003] However, existing methods for constructing brain networks and locating key nodes based on PET-MR images still have the following shortcomings: First, most existing brain network analyses adopt standardized static population analysis processes, which are difficult to characterize the dynamic evolution characteristics of brain networks. Moreover, dynamic network construction methods based on fixed sliding windows are easily affected by window parameters, resulting in insufficient stability and accuracy. Second, PET and MR images usually adopt a simple fusion approach after independent feature extraction, lacking in-depth cross-modal feature interaction and failing to fully leverage the advantages of multimodal complementarity. Third, existing localization methods generally rely on standard brain template registration, which easily weakens individual differences in brain structure, while directly regressing three-dimensional coordinates lacks brain anatomical constraints, affecting the accuracy of key node localization. Finally, existing deep learning models lack interpretability, making it difficult to explain the criteria for determining key nodes, thus limiting their practical application in brain science research and analysis.

[0004] Therefore, it is necessary to design a spatial localization method for key nodes of cognitive networks that can achieve deep PET-MR multimodal fusion, dynamic brain network construction, precise localization of individualized key nodes, and good interpretability. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by proposing a spatial localization method for key nodes in cognitive networks based on PET-MR. The method aims to solve problems such as insufficient multimodal fusion, lack of adaptive mechanisms in dynamic network modeling, insufficient spatial localization accuracy, and a single method for evaluating the reliability of results. The method includes the following steps:

[0006] Step S1: Multimodal Image Data Acquisition and Preprocessing. Acquire integrated PET-MR simultaneous scan data of the subjects, including PET metabolic images and MR structural images; perform standardized preprocessing on the PET metabolic images and MR structural images respectively; perform brain region segmentation based on individual MR structural images, define a set of brain region nodes, and define the initial spatial coordinates of each brain region node as its anatomical centroid coordinates; define the cognitive network as a weighted graph with brain region nodes as vertices and metabolic covariance or structural connectivity between brain regions as edges; define key nodes as brain region nodes with the highest network centrality score in the cognitive network.

[0007] Step S2: Semantic Prototype-Guided Cross-Modal Feature Alignment and Fusion. A medical semantic encoder is constructed to encode the functional semantic description text of each brain region into semantic prototype vectors; a cross-modal feature extraction network is constructed to extract multi-scale metabolic feature maps and multi-scale anatomical feature maps from PET metabolic images and MR structural images, respectively; preliminary fusion features are generated through a cross-modal cross-attention mechanism; guided by the semantic prototype vectors, the alignment loss between the fusion features of each brain region and the corresponding semantic prototype is calculated, and semantically enhanced cross-modal fusion feature maps are generated by minimizing this loss; based on the cross-modal fusion feature maps, guided super-resolution reconstruction of PET metabolic images is performed using MR structural information to generate high-resolution metabolic maps, and marginal space confidence maps are output simultaneously.

[0008] Step S3: Construction of a neural modulation-driven dynamic cognitive network. Based on high-resolution metabolic maps, time-series metabolic data of each brain region are extracted; a global neural modulator is constructed, which outputs the dynamic modulation gain coefficients of each brain region node and each network edge according to the whole-brain metabolic state of the current time window and the hidden state of the previous time window; the basal metabolic cocorrelation matrix is ​​calculated using the sliding window covariance analysis method, and the dynamic modulation gain coefficients are applied to the basal metabolic cocorrelation matrix in a multiplicative manner to generate a dynamic metabolic network edge weight sequence; spatial proximity decay weights are introduced as prior constraints and embedded into the dynamic edge weight calculation; a structural connectivity network is constructed based on MR structural images; the dynamic metabolic network edge weight sequence, the static metabolic network, and the structural connectivity network are multiplexed and fused to generate a multiplexed cognitive network.

[0009] Step S4: Hierarchical Community Detection and Spatial Guidance Field-Driven Node Coordinate Offset. Hierarchical community detection is performed on the multiplexed cognitive network, identifying functional community structures at each level. For each brain region node, its cross-level influence score in different levels of communities is calculated. The importance of all brain region nodes is ranked based on the combined influence scores at each level, outputting a candidate set of key nodes. A spatial guidance field is constructed based on the candidate set of key nodes, with each candidate node exerting a spatial guidance effect on its surrounding brain regions. The total spatial guidance vector received by each brain region node is calculated, and a first-level spatial offset correction is applied to the anatomical centroid coordinates defined in Step S1 to obtain the spatial coordinates corrected by the guidance field.

[0010] Step S5: Spatial Residual Regression and Coordinate Refinement for Differentiable Anatomical Constraints. A spatial residual regression graph neural network is constructed, using a multiplexed cognitive network and spatial coordinates corrected by the guiding field as input to predict the spatial coordinate offset residuals of each brain region node. The training loss of the graph neural network includes coordinate offset regression loss and geometric consistency loss. The geometric consistency loss is used to constrain the spatial offset directions of adjacent brain region nodes with functional coupling to tend to be consistent. The marginal spatial confidence map is sampled to each brain region node as sample weights for coordinate regression. The graph neural network outputs information explaining the contribution of the node criticality score through a graph attention mechanism. Candidate key nodes are selected based on the criticality score output by the graph neural network, and the predicted spatial coordinate offset residuals are superimposed on the coordinates corrected by the guiding field.

[0011] Step S6: Refining the spatial coordinates of differentiable anatomical constraints. The spatial coordinate offset residuals predicted in Step S5 are parameterized as deformation metric fields of local brain regions. A differentiable anatomical constraint solver is constructed, using the anatomical centroid coordinates in Step S1 as the reference configuration and the predicted deformation metric field as the target metric. The optimal spatial coordinates that satisfy anatomical rationality are solved through differentiable constraint optimization. The constraint optimization includes at least the non-penetrating constraint of the brain sulcus and gyrus boundaries, the smooth manifold regularization constraint, and the anatomical label consistency constraint. The refined coordinates output by the solver are used as the final positioning coordinates.

[0012] Step S7: Multi-granularity confidence assessment and localization error feedback. The Bootstrap resampling method is used to robustly assess the final localization coordinates, calculate the confidence intervals for each key node coordinate, and construct confidence ellipsoid parameters. A cross-time window localization stability index is calculated to quantify the sensitivity of the localization result to time window selection. An anatomical-functional semantic validator is constructed, comparing the localization result with the brain anatomy knowledge graph and functional cognitive graph, and outputting a semantic reasonableness score. The confidence ellipsoid volume is used as a localization error feedback signal to inversely adjust the boundary constraint strength of the super-resolution reconstruction network in step S2, forming a closed-loop optimization. The localization result and multi-granularity confidence information are output in a standardized format.

[0013] By adopting the above solution, the beneficial effects achieved by the present invention are as follows:

[0014] This invention employs a semantic prototype-guided cross-modal feature alignment and fusion mechanism. It utilizes a medical semantic encoder to encode the functional semantic description text of each brain region into semantic prototype vectors. These semantic prototype vectors serve as supervisory signals to guide cross-modal cross-attention interaction and alignment of multimodal features, causing functionally similar brain regions to cluster together and functionally different brain regions to separate in the fused feature space. Simultaneously, it utilizes MR structural information to perform guided super-resolution reconstruction of PET metabolic images, generating high-resolution metabolic maps and outputting marginal space confidence maps, thereby improving the anatomical-functional consistency of cross-modal feature fusion.

[0015] This invention constructs a global neural modulator that adaptively generates dynamic modulation gain coefficients for each brain region node and network edge based on the whole-brain metabolic state. These coefficients are applied multiplicatively to the sliding window co-correlation matrix, enabling the dynamic network edge weights to respond to temporal changes in brain state and overcoming the influence of fixed window parameters on network estimation stability. Furthermore, it identifies multi-scale functional community structures through hierarchical community detection, calculates cross-level influence scores to screen key node candidate sets, and uses the spatial guidance field of candidate nodes to guide coarse coordinate localization. Finally, it refines the coordinates by regressing the spatial residuals using a spatial residual regression graph neural network and combining it with a differentiable anatomical constraint solver that includes boundary impenetrability constraints, smooth manifold regularization constraints, and anatomical label consistency constraints. This achieves individualized spatial localization and avoids invalid anatomical localization.

[0016] This invention outputs the contribution explanation information of the key scores of each node through a graph attention mechanism, and uses Bootstrap resampling to calculate the confidence ellipsoid parameters, cross-time window localization stability index, and anatomical-functional semantic verifier score to provide a confidence assessment of the localization results from multiple dimensions. The confidence ellipsoid volume is used as a feedback signal to adjust the boundary constraint strength of the super-resolution reconstruction network, forming a closed loop from localization uncertainty to image reconstruction. At the same time, the localization results and multi-granularity confidence information are output in a standardized format. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the overall process of a spatial localization method for key nodes in a cognitive network based on PET-MR proposed in this invention.

[0018] Figure 2 This is a schematic diagram of the structure of the cross-modal feature fusion and dynamic cognitive network construction of the present invention;

[0019] Figure 3 This is a schematic diagram of the process for refining the spatial coordinates of key nodes and evaluating confidence levels in this invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0021] All integrated PET-MR image data used in the embodiments of this invention have undergone anonymization processing, removing subject identification identifiers and retaining only image pixel data. Data processing complies with relevant laws and regulations concerning personal information protection and data security. This invention is a medical image signal processing method and does not perform disease diagnosis on the human body or output spatial localization parameters of key nodes in the brain's cognitive network; it is intended solely for scientific research analysis or physician reference.

[0022] Example 1:

[0023] This embodiment uses multimodal brain imaging data acquired by an integrated PET-MR device as an example to illustrate a spatial localization method for key nodes of cognitive networks based on PET-MR proposed in this invention.

[0024] like Figure 1 As shown, the method in this embodiment includes the following steps:

[0025] Step S1: Multimodal image data acquisition and preprocessing

[0026] Acquire integrated PET-MR simultaneous scan data from subjects, including PET metabolic images and MR structural images.

[0027] First, image reconstruction, attenuation correction, scattering correction, and spatial normalization were performed on PET metabolic images; offset field correction, cranial removal, and spatial registration were performed on MR structural images.

[0028] Brain region segmentation is performed based on individual MR structural images, dividing the brain into multiple anatomical brain regions. Each brain region serves as a node in the cognitive network, and the anatomical centroid of the corresponding brain region is used as the initial three-dimensional spatial coordinates.

[0029] Metabolic features of each brain region were extracted from PET metabolic images, and spatial adjacency and structural correlation features between brain regions were extracted from MR structural images. This process constructed structural relationships between brain regions, forming a cognitive network with brain region nodes as vertices and inter-brain region relationships as edges. The importance of brain region nodes was determined by their overall influence within the cognitive network.

[0030] Step S2: Semantic Prototype-Guided Cross-Modal Feature Alignment and Fusion

[0031] like Figure 2As shown, the PET metabolic images and MR structural images obtained in step S1 are input into the cross-modal feature extraction network to obtain PET metabolic features and MR structural features.

[0032] For each brain region node, the corresponding medical semantic description information is obtained, and the semantic encoding module generates a brain region semantic prototype vector in d-dimensional real space.

[0033] The cross-modal cross-attention mechanism is used to realize the information interaction between PET features and MR features. Specifically, the query matrix obtained by linear transformation of PET features is used to perform standard Transformer cross-attention calculation with the key matrix and value matrix obtained by linear transformation of MR structural features to obtain PET-MR cross-modal fusion features.

[0034] Furthermore, using the semantic prototype vectors of brain regions as constraints, semantic consistency optimization is performed on the fused features, and the semantic constraint loss function is defined as follows:

[0035] ;

[0036] in, This represents the total number of nodes in the brain region. and All are brain region node indexes. , and ; and These are the fusion feature vectors of the v-th and u-th brain region nodes, respectively; This is the semantic prototype vector corresponding to the v-th brain region node; represents the L2 norm of a vector; · represents the dot product operation of vectors; To separate the loss weight coefficients and control the discriminative strength between features of different brain regions; A similarity threshold is used to determine whether features from different brain regions need to be separated from each other; The operation is performed to maximize the value; the first term is used to improve the consistency between the fused features and the semantic prototypes of the corresponding brain regions; the second term is used to enhance the feature discrimination ability between different brain regions.

[0037] Based on the obtained fusion features, guided super-resolution reconstruction of PET metabolic images is performed using MR structural information to generate high-resolution metabolic maps. Simultaneously, a dual-branch output is configured in the super-resolution reconstruction network: one branch outputs the reconstructed high-resolution PET metabolic map, and the other outputs a 3D edge space confidence map with values ​​between 0 and 1. The edge space confidence map represents the reliability of brain region boundaries at different spatial locations and serves as weight information in the subsequent spatial localization optimization process.

[0038] Step S3: Construction of a neural modulation-driven dynamic cognitive network

[0039] Based on the high-resolution PET metabolic map obtained in step S2, dynamic metabolic time series of each brain region node were extracted.

[0040] A neural modulation module is constructed to generate node modulation coefficients and edge connection modulation coefficients (i.e., the dynamic modulation gain coefficients) based on the whole brain metabolic state and historical state information within the current time window.

[0041] The node modulation coefficients are expressed as:

[0042] ;

[0043] The side-connected modulation coefficients are expressed as:

[0044] ;

[0045] in, This represents the global brain state encoding vector, which is obtained by global pooling of the metabolic features of all brain regions within the current time window. and These represent the multilayer perceptron mapping functions in the node modulator and the edge modulator, respectively; Represents the hyperbolic tangent activation function; This represents the node modulation coefficient, which scales the contribution of the metabolic activity of a single brain region node to the network edge weights. This represents the edge connection modulation coefficient, which directly regulates the connection strength between brain regions.

[0046] The basal metabolic cocorrelation matrix between brain regions was calculated using the Pearson correlation coefficient method with a sliding time window, thus obtaining the degree of metabolic correlation between brain region i and brain region j within the k-th time window. The basic cocorrelation matrix is ​​dynamically adjusted using neural modulation coefficients, and the adjusted edge weights are calculated using the following formula:

[0047] ;

[0048] in, This represents the edge weights of the metabolic network after dynamic modulation; This represents the connection modulation coefficient between brain region i and brain region j within the k-th time window; and represents the node modulation coefficients of brain region i and brain region j within the k-th time window, respectively; the first term is used to adjust the changes in the connectivity between brain regions, and the second term is used to adjust the local contribution weight according to the metabolic state of the node itself.

[0049] Furthermore, a spatial proximity decay weight based on a Gaussian kernel function is introduced as a prior constraint, and the spatial proximity decay weight is calculated based on the Euclidean spatial distance between the anatomical centroids of two brain regions. The final formula for calculating the edge weights of the dynamic network is as follows:

[0050] ;

[0051] in, Indicates the first Brain regions within a time window With brain regions The final dynamic network edge weights between them; The weight coefficient representing the spatial proximity decay weight adjusts the contribution of spatial prior information in dynamic network construction.

[0052] By fusing dynamic metabolic networks, static metabolic networks, and structural association networks constructed based on MR structural features, a multiplexed cognitive network is obtained.

[0053] Step S4: Hierarchical community detection and spatially guided field-driven node coordinate offset

[0054] like Figure 3 As shown, hierarchical community detection is performed on the multiplexed cognitive network constructed in step S3 to identify brain functional community structures at different scales.

[0055] In each level of the community structure, network centrality indices of brain region nodes are calculated, including degree centrality, betweenness centrality, and eigenvector centrality. Combined with the tensor eigenvalue centrality of nodes in the multiplexing network, the cross-level influence score of nodes is calculated. The calculation formula is as follows:

[0056] ;

[0057] in, brain region nodes Cross-level influence score; Indexing brain region nodes, ; For community testing hierarchy index, These correspond to four levels, from coarse-grained to fine-grained. For the first The number of centrality indicators used at each level; For indexes of centrality indicator types, ; For nodes In the The first in the hierarchy Class centrality index value; For the first The first in the hierarchy The weighting coefficients of the class centrality index; For nodes Tensor feature centrality in multiplexed networks; These are the tensor eigencenter weighting coefficients.

[0058] Nodes in all brain regions are sorted by their influence scores, and high-influence nodes are selected as a candidate set of key nodes based on a preset ratio or influence threshold. .

[0059] Furthermore, a spatial guidance field is constructed based on the candidate set of key nodes. For any brain region node... The formula for calculating the spatial guiding vector of all candidate key nodes is as follows:

[0060] ;

[0061] in, For nodes The total spatial guiding vector received; Index for candidate key nodes. ; Candidate key nodes The cross-level influence score is used as the strength coefficient of the guiding role of the node; and They are nodes and candidate key nodes Current spatial coordinate vector; Let L2 be the norm of the vector; To prevent stable parameters with zero denominators, a small positive constant is chosen.

[0062] Nodes are calculated based on the spatial guiding vector and preset step size parameters. The first-level spatial coordinate offset is then superimposed onto the node. The spatial coordinates obtained from the current spatial coordinates are corrected by the spatial guiding field. Through the above-described spatial guidance field-driven processing, more reasonable initial spatial positioning values ​​are obtained for potential critical node regions.

[0063] Step S5: Spatial residual regression and coordinate refinement with differentiable anatomical constraints

[0064] The multiplexed cognitive network obtained in step S3 and the spatially corrected coordinate input spatial residual regression graph neural network obtained in step S4 are combined.

[0065] The spatial residual regression neural network predicts the three-dimensional spatial offset residual vector of each brain region node based on the network connectivity and spatial location information between brain regions. During model optimization, localization error constraints and spatial consistency constraints are introduced.

[0066] The formula for calculating the coordinate offset regression loss is:

[0067] ;

[0068] in, This represents the coordinate offset regression loss; Indicates the total number of nodes in the brain region; Indicates the index of brain region nodes. ; This indicates that the confidence map of the marginal space was sampled up to the first... The weight coefficients of each brain region node are used to reduce the impact of low-confidence region localization errors on model training; The network prediction of the first Spatial offset residual vector of each node; Indicates the first The reference space offset constraint of each node can be constructed using standard brain templates, expert annotation results, or historical positioning results. This represents the square of the L2 norm of a vector.

[0069] Furthermore, spatial consistency constraints are introduced between strongly connected brain regions, and the geometric consistency loss is defined as follows:

[0070] ;

[0071] in, This represents the geometric consistency loss; Indicates the index of brain region node pairs; The set of strongly connected brain regions is defined as the network edge weight. The set of node pairs that exceeds a preset threshold; and These represent the network predictions of the first, second, and third generations, respectively. The node and the first Spatial offset residual vector of each node; This indicates that the first step obtained in step S3 is... The node and the first The final network edge weights between nodes are used to weight the constraint strength.

[0072] The overall objective function of the spatial residual regression graph neural network is:

[0073] ;

[0074] in, This represents the total training loss of the spatial residual regression graph neural network. represents the weighting coefficients of the spatial consistency constraint, balancing the relative contributions of regression loss and consistency loss to the total loss.

[0075] The spatial offset residual vector of the i-th node predicted by the network The coordinates of node i are superimposed onto the coordinates obtained in step S4 after spatial guidance field correction, resulting in the spatial coordinates of node i refined by the spatial residual regression graph neural network.

[0076] Meanwhile, the graph attention mechanism is used to output the key contribution information of each node, including the attention weights of the neighboring nodes that make the greatest contribution to the node and their connecting edges, as well as the metabolic feature channel and structural feature channel that make the greatest contribution, which are used to explain the basis for determining key nodes.

[0077] Step S6: Refining the spatial coordinates of the differentiable anatomical constraints

[0078] The spatial offset results predicted in step S5 are converted into a measure of local brain region deformation.

[0079] Using the initial anatomical space coordinates obtained in step S1 as a reference configuration, a differentiable anatomical constraint solver is constructed, and the predicted coordinates are further refined by optimizing the following objective function:

[0080] ;

[0081] in, The overall optimization objective function is the differentiable anatomical constraint solver. To measure the constraint loss, the distance between the brain region functional features corresponding to the constrained prediction spatial coordinates and the target functional features is used. The boundaries of the brain sulci and gyri are impenetrable constraints, ensuring that the refined node positions do not deviate from the corresponding anatomical boundaries of the brain regions. To ensure spatial manifold smoothness constraints and maintain the geometric continuity of the cortical surface; Constraints were imposed on the consistency of brain region anatomical labels to ensure that the refined voxel classification was consistent with the MR segmentation results; These are the weight coefficients for the corresponding constraint terms, balancing the relative contributions of each constraint to the overall objective.

[0082] Through iterative optimization, the final spatial coordinates simultaneously satisfy the functional network localization results and the constraints of brain tissue anatomical structure. After optimization, the final spatial coordinates of the key nodes are obtained. This serves as the location result for key nodes.

[0083] Step S7: Multi-granularity confidence assessment and positioning error feedback

[0084] The stability of the final positioning result obtained in step S6 is evaluated.

[0085] A repeated perturbation analysis method was employed to sample the positioning results multiple times and repeatedly execute the key node positioning process. The spatial coordinate distribution of key nodes under different perturbation conditions was statistically analyzed to obtain the positioning confidence interval. Simultaneously, the changes in the ranking of key nodes under different time windows were calculated to evaluate the sensitivity of the positioning results to changes in the time window.

[0086] Furthermore, an anatomy-functional semantic verifier is constructed, and the localization results are compared with the brain anatomy knowledge graph and cognitive function graph to obtain a semantic rationality score.

[0087] The boundary constraint strength in the super-resolution reconstruction process in step S2, based on the positioning error feedback, is updated using the following formula:

[0088] ;

[0089] in, The updated boundary constraint strength parameters; The current boundary constraint strength parameter; The feedback adjustment coefficient controls the response amplitude of the confidence ellipsoid volume deviation to the constraint strength adjustment. The confidence ellipsoid volume corresponding to the current positioning result reflects the overall level of positioning uncertainty; The reference confidence ellipsoid volume is predetermined using historical data or population statistics.

[0090] Through the aforementioned feedback mechanism, a closed-loop adjustment is achieved from localization error assessment to image reconstruction optimization. The final output includes: key brain region names, individual spatial three-dimensional coordinates, network importance score, 95% confidence interval, and semantic plausibility score.

[0091] This embodiment reduces localization bias caused by low spatial resolution of PET, inconsistency of cross-modal information, and insufficient brain region boundary constraints by using PET-MR multimodal fusion, dynamic cognitive network construction, spatial guided field localization, graph network residual optimization, and anatomical constraint refinement. This improves the accuracy, stability, and interpretability of spatial localization of key nodes.

[0092] Example 2:

[0093] This embodiment illustrates the application of the method of the present invention in the quantitative scientific research analysis of brain cognitive networks related to neurodegenerative diseases.

[0094] The specific process for processing subject PET-MR data using this method is as follows:

[0095] First, the subject's PET-MR data was acquired and preprocessed according to step S1 of Example 1. The MR structural images were segmented into brain regions, 90 brain region nodes were defined, and a cognitive network with brain region nodes as vertices was established.

[0096] Secondly, steps S2 to S6 are performed to perform cross-modal feature fusion, dynamic cognitive network construction, and key node spatial localization on PET-MR multimodal image data of different subjects; PET metabolic information and MR structural information are fused to construct a multi-path multiplexing cognitive network corresponding to each subject, and key nodes and their spatial coordinates are determined according to the node influence, network connection relationship and spatial constraint optimization process.

[0097] When analyzing different groups of subjects, some abnormal states of cognitive networks can be manifested as changes in the connection strength within the default mode network, and the spatial distribution of key nodes shows regional characteristics related to abnormal changes in cognitive networks. Under normal cognitive conditions, key nodes can be distributed in multiple functional systems such as the fronto-parietal network, the default mode network, and the salience network.

[0098] Finally, step S7 is executed to evaluate the reliability of the localization results and output a standardized localization report. The localization report includes the names of key nodes, the individual's three-dimensional spatial coordinates, the 95% confidence interval, and the semantic reasonableness score, providing a quantitative reference for quantitative research on brain cognitive networks.

[0099] Example 3:

[0100] This embodiment illustrates the engineering implementation of the method of the present invention in a PET-MR image analysis platform.

[0101] This method can be deployed on medical image processing servers equipped with GPUs or heterogeneous computing capabilities.

[0102] Computational resource configuration: The cross-modal feature extraction network and super-resolution reconstruction network in step S2 are accelerated by GPU; the dynamic network calculation, graph network inference and spatial optimization processes in steps S3 to S6 can be completed by GPU, CPU or heterogeneous computing according to the computing resource configuration to meet the processing requirements in the medical image analysis process.

[0103] Integration with medical image processing workflow: After the integrated PET-MR device completes the scan, the DICOM image data is automatically transmitted to the processing workstation, triggering the execution flow of this method. The system sequentially completes image preprocessing, brain region segmentation, multimodal feature fusion, cognitive network construction, and spatial localization of key nodes, ultimately generating a standardized analysis report, which can be sent back to the medical image management system.

[0104] The system provides a mechanism to handle abnormal situations: When the semantic rationality score output by the anatomical-functional semantic verifier is lower than the preset threshold, the system marks the current positioning result as abnormal and prompts for manual review; when the confidence ellipsoid volume corresponding to the positioning result exceeds the preset reference range, the system adjusts the boundary constraint strength in step S2 according to the feedback information and re-executes the positioning optimization process to improve the stability of the positioning result.

[0105] Model training and operation methods:

[0106] The neural network modules involved in this method include a cross-modal feature extraction network, a super-resolution reconstruction network, a global neural modulator, and a spatial residual regression graph neural network. The parameters of these network modules can be obtained through pre-training and end-to-end optimization. The training process uses PET-MR data samples containing reference localization results for optimization.

[0107] In actual operation, after new PET-MR image data is input, the system sequentially performs image processing, cognitive network construction, key node localization and confidence assessment according to steps S1 to S7. The network parameters remain fixed, and only forward computation is performed. Finally, individualized key node spatial localization results are output.

[0108] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. Those skilled in the art, inspired by this description, may design similar structures and embodiments without departing from the spirit of the invention, and such designs should all fall within the protection scope of the present invention.

Claims

1. A spatial localization method for key nodes in a cognitive network based on PET-MR, characterized in that, Includes the following steps: Step S1: Acquire PET metabolic images and MR structural images from integrated PET-MR simultaneous scanning; perform brain region segmentation based on individual MR structural images, using each brain region as a node and the anatomical centroid as the initial spatial coordinates; Step S2: Construct a functional semantic prototype of brain regions, perform cross-modal feature fusion on PET metabolic images and MR structural images and align features with the semantic prototype; use MR structural information to perform guided super-resolution reconstruction of PET metabolic images to generate high-resolution metabolic maps. Step S3: Extract node time series based on high-resolution metabolic map and generate dynamic edge weight sequence; extract static metabolic network based on PET metabolic image; construct structural connectivity network based on MR structural image; fuse dynamic edge weight sequence, static metabolic network and structural connectivity network to form multiplexed cognitive network; Step S4: Perform hierarchical community detection on the multi-path multiplexing cognitive network, calculate cross-level influence scores by integrating multi-level centrality indicators, and screen key node candidate sets; use the candidate sets to construct a spatial guidance field to correct the initial spatial coordinates and obtain the corrected coordinates; Step S5: Using the multiplexed cognitive network and the corrected coordinates as input, predict the node offset residuals through the spatial residual regression graph neural network trained with geometric consistency constraints, and superimpose them onto the corrected coordinates to obtain the refined coordinates; Step S6: Convert the offset residuals into a local deformation metric field, optimize the refined coordinates using a differentiable dissection constraint solver, and output the final refined coordinates; Step S7: Evaluate the confidence level of the final refined coordinates, use the confidence ellipsoid volume as a feedback signal to adjust the boundary constraint strength of the super-resolution reconstruction, and output the standardized positioning results.

2. The method according to claim 1, characterized in that, In step S2, feature alignment is performed using semantic prototypes as supervision signals. Specifically, this includes: constructing a semantic alignment loss to make the fused features of each node closer to the corresponding semantic prototype; simultaneously constructing a discrimination loss to apply a discrimination penalty when the similarity between the fused features of different nodes exceeds a preset threshold, thereby enhancing the feature discrimination between different nodes; and the guided super-resolution reconstruction also generates a marginal space confidence map to characterize the reliability of brain region boundaries.

3. The method according to claim 1, characterized in that, In step S3, the specific method for generating the dynamic edge weight sequence is as follows: calculate the basic cocorrelation matrix based on the node time series; construct a global neural modulator to generate node dynamic gain coefficients and edge dynamic gain coefficients according to the global brain state encoding of the current time window; modulate the basic cocorrelation matrix using the node dynamic gain coefficients and edge dynamic gain coefficients, and obtain the dynamic edge weight sequence by combining spatial proximity priors; extract the static metabolic network based on PET metabolic images; and the multiplexed cognitive network is formed by fusing the dynamic edge weight sequence, the static metabolic network, and the structural connectivity network.

4. The method according to claim 1, characterized in that, In step S4, the multi-level centrality index includes degree centrality, betweenness centrality and eigenvector centrality. The calculation of cross-level influence score also incorporates the tensor eigenvalue centrality of each node in the multiplexed cognitive network. The specific method for constructing the spatial guidance field is as follows: each candidate node is used as an intensity coefficient based on its influence score, and the coefficients are superimposed along the spatial direction determined by the difference between the candidate node coordinates and the target node coordinates to obtain the total guidance vector acting on the target node, and the initial spatial coordinates are corrected accordingly.

5. The method according to claim 2, characterized in that, In step S5, when predicting node offset residuals using a spatial residual regression graph neural network, a regression loss is constructed using the edge spatial confidence map as weights, and a weighted constraint is applied to the error term between the predicted offset residuals and the reference offset. The geometric consistency constraint targets strongly connected node pairs whose network connection strength exceeds a preset threshold, constraining the difference in their offset residuals to be minimized, and using the corresponding network edge weights as the strength weights of this constraint. The total loss of the graph neural network is the weighted sum of the regression loss and the geometric consistency loss. The graph neural network also outputs explanatory information on the critical contribution of nodes.

6. The method according to claim 1, characterized in that, In step S6, the differentiable anatomical constraint solver includes metric constraints, boundary impenetrability constraints, manifold smoothing constraints, and anatomical label consistency constraints. Among them, the metric constraints are used to constrain the feature distance between the brain region functional features corresponding to the predicted spatial coordinates and the preset target functional features; the boundary impenetrability constraints are used to prevent the optimized node positions from deviating from the anatomical boundaries of the corresponding brain regions; the manifold smoothing constraints are used to maintain the geometric continuity of the cortical surface; and the anatomical label consistency constraints are used to ensure that the classification results of the optimized voxels are consistent with the segmentation results of the MR structural images.

7. The method according to claim 1, characterized in that, In step S7, the specific method for adjusting the boundary constraint strength of the super-resolution reconstruction is as follows: when the confidence ellipsoid volume is larger than the preset reference volume, the boundary constraint strength is increased according to the deviation ratio; when the confidence ellipsoid volume is smaller than the preset reference volume, the boundary constraint strength is decreased according to the deviation ratio. The adjustment process is equipped with a feedback adjustment coefficient to control the response amplitude.

8. The method according to claim 7, characterized in that, In step S7, when the deviation between the confidence ellipsoid volume and the preset reference volume exceeds the preset tolerance threshold, the boundary constraint strength in step S2 is adjusted, forming a closed-loop iteration of positioning evaluation and image reconstruction. When the deviation does not exceed the preset tolerance threshold, the current localization parameters are maintained and the current localization result is output; the standardized localization result includes at least the name of the key brain region, the individual's three-dimensional spatial coordinates, the network importance score, the confidence interval, and the semantic reasonableness score.

Citation Information

Patent Citations

  • Cross-modal fusion attention mechanism Alzheimer disease early risk assessment method and system based on PET image, MR image and cognitive scoring data

    CN122599024A

  • Generating brain network impact scores

    US11264138B1