CN-117153340-B - Method for self-adaptive weighted fusion based on multi-functional connection network

Abstract

The invention discloses a method based on self-adaptive weighted fusion of multiple functional connection networks, which comprises the following steps of preprocessing brain functional image data output by a functional magnetic resonance instrument, dividing the preprocessed brain functional image into a plurality of brain regions according to a standard template, calculating connection weights among the brain regions based on multiple correlations, respectively constructing the functional connection networks, and self-adaptively learning the fusion weights for each functional connection network by using supervision information so as to obtain the functional connection network containing multiple interaction information. The method uses simple operation to fuse interaction information of different visual angles, can judge the importance of each functional connection network in the fusion process, and integrates the subsequent feature selection and classification into one model, thereby having important theoretical significance and practical application value.

Inventors

• ZHANG LIMEI
• ZHANG CHAOJUN
• Qiao Lishan

Assignees

• 山东建筑大学
• 聊城大学

Dates

Publication Date

20260505

Application Date

20230228

Claims (2)

1. 1. The method for self-adaptive weighted fusion of the network based on the multi-functional connection is characterized by comprising the following steps: (1) The method comprises the following steps of preprocessing brain functional images acquired by functional magnetic resonance equipment by using a DPARSF toolbox of matlab, namely removing the first p time points of the magnetic resonance images, performing head movement and time layer correction on the images, removing influences generated by ventricular, white matter signals and head movement higher-order effects, registering the corrected images to a standard space, and performing band-pass filtering on the attempted images by using 0.01-0.1HZ time to reduce the influences of heart beat and respiration; (2) After preprocessing brain functional image data, dividing the brain into a plurality of brain areas based on an automatic anatomical labeling map, and extracting an average time sequence of each brain area to be expressed as Wherein Represent the first A time series of individual brain regions, wherein, The length of the time series is indicated, Representing the number of brain regions; (3) Calculating multiple correlations between every two brain regions based on the average time sequence of the brain regions to create multiple functional connection networks W, wherein the functional connection networks are expressed as symmetrical matrixes mathematically, and in order to avoid redundancy of data, the upper triangle characteristic of the functional connection networks is selected and leveled to form a row vector; (4) On the basis of the created multiple functional connection networks, each functional connection network self-adaptive learning fusion weight by using supervision information Thereby obtaining a functional connection network containing multiple correlations ; (5) The model of the adaptive weighted fusion of the various correlation function connection networks is as follows: (1) (2) Wherein, the Represent the first To be tested The functions of the different correlation estimates connect to the network, Representation of C is the weight vector in the L 1 -norm SVM; Is the first M is the number of subjects, K represents the number of functional connection networks; is an error penalty parameter that is used to determine the error penalty, Is a regularization parameter; the weight vector C tends to be sparse, Preventing model degradation to Constraint(s) ; (6) The model in the step (5) involves Two variables, alternate optimization was used to solve the model: First, fix Updating C, the model can be reduced to the following objective function L 1 -norm SVM: (3) The L 1 -norm SVM is solved using liblinear packages in Matlab; second, in the case of fixed C, then update The model can be simplified as: (4) (5) introducing relaxation variables The above models (4), (5) can be rewritten as: (6) (7) Using joint variables Solving a quadratic programming problem, and solving by using quadprog functions in Matlab; (7) When the difference between the function value of the next iteration and the function value of the previous iteration in the step (5) is smaller than a determined threshold And (5) the algorithm converges, the iteration is stopped, the optimal function value is output, and otherwise, the calculation is repeated.
2. 2. The method of claim 1, wherein the step (3) calculates a plurality of correlations between two brains to create a plurality of functionally connected networks W, wherein the plurality of correlations includes a full correlation, a partial correlation, a mutual information, and a correlated correlation.

Description

Method for self-adaptive weighted fusion based on multi-functional connection network Technical Field The invention relates to the technical field of image processing, in particular to a method for self-adaptive weighted fusion based on a multi-functional connection network. Background Autism spectrum disorder is a childhood neurodevelopmental disorder that manifests itself primarily as social disorders, language communication disorders, shortness of interest, and palpation. At present, the diagnosis of autism mainly depends on simple symptom observation and experience of clinicians, and is easy to cause high misjudgment rate, so that treatment is delayed. Thus, the search for a reliable, automated autism-assisted diagnostic strategy, especially in its early stages, is of increasing interest in the fields of psychiatry and neuroscience. Resting state functional magnetic resonance imaging is a technique that measures blood oxygen level dependent signals of a subject without performing any specific task and has been widely used for early diagnosis of autism. Since autism spectrum disorders were found to tend to disrupt the neural connection of different brain regions, constructing a high quality functional connection network becomes a key issue in capturing subtle abnormal changes in brain regions of autistic subjects. From a mathematical perspective, each node in the functional connection network of the brain corresponds to a brain region of interest, and each side corresponds to an interaction relationship between two brain regions. To estimate edges in a functionally connected network, researchers have developed many different approaches including pearson correlation coefficient (PC), sparse Representation (SR), mutual Information (MI), and related correlations (CC), among others. Based on these approaches, functionally connected networks can be well estimated, but each of them can only capture a single type of relationship between different brain regions, and thus it is difficult to model complex interactions in the brain. Inspired by the concept of multi-view learning, fusing complementary features of multiple views is a popular approach. For example, CN115099369A discloses a network fusion method based on functional connection and structural connection, which has the technical contents that a cerebral cortex is divided into different brain areas and networks based on an AAL template and a Yeo brain network template, fMRI and DTI data are preprocessed, a large-scale brain connection matrix is constructed on the preprocessed fMRI and DTI data, and network horizontal weighting probability between functional-structural connection is calculated. The network fusion method has the technical problems that based on the network horizontal weighting probability, the classification effect only depends on the quality of the constructed functional connection and the structure connection network, the difference between different tested is not considered, and the important role played by the supervision information in the fusion process is ignored. And as a Chinese patent of publication No. CN112418337B, a multi-feature fusion magnetic resonance image data classification method based on brain function super-network model is disclosed, a brain function super-network is constructed by utilizing a composition MCP method, then the features of brain areas are extracted by utilizing various different indexes to comprehensively quantify the topology of the brain function super-network, and finally the method is used for diagnosing brain diseases. When the method is used for multi-feature fusion, KS test is utilized, feature distribution is utilized to split fusion and classification, and the follow-up classification can not influence the previous fusion, namely, the previous fusion is good or bad, the fused features can not be reflected, and the true brain feature condition can not be better reflected. Disclosure of Invention The invention aims to provide a method for self-adaptive weighted fusion of a plurality of functional connection networks, which utilizes supervision information to guide the fusion of the functional connection networks, automatically learns the weight of each functional connection network, can better solve the problem of the fusion weight, fuses complementary information of all visual angles, namely, information of full correlation, partial correlation, mutual information and related equal angles of the functional connection networks, integrates brain network fusion and subsequent classification into one model, leads the prior fusion weight learning to the classification result, and can continuously update the learning when the learned fusion weight is not good, so that the learned fused brain network can more truly reflect the characteristics of the brain. In order to solve the technical problems, the invention adopts the following technical means: A method for self-adaptive weighted fusion based on a