A method for adaptive weighted fusion based on multiple functional connection networks

Abstract

The application discloses a method for adaptive weighted fusion based on multiple functional connection networks, and steps are as follows: pre-processing brain function image data output by a functional magnetic resonance instrument; dividing the pre-processed brain function image into a plurality of brain regions according to a standard template; calculating connection weights between each brain region based on multiple correlations, and respectively constructing functional connection networks; and adaptively learning and fusing weights for each functional connection network by using supervision information, so that a functional connection network containing multiple interaction information is obtained. The method has important theoretical significance and practical application value, and can fuse interaction information from different perspectives by using simple operation, can judge importance of each functional connection network in the fusion process, and integrates subsequent feature selection and classification into a model.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically a method based on adaptive weighted fusion of multiple functional connection networks. Background Technology

[0002] Autism spectrum disorder is a neurodevelopmental disorder in children, primarily characterized by social impairments, language and communication difficulties, restricted interests, and repetitive behaviors. Currently, the diagnosis of autism relies mainly on simple symptom observation and the experience of clinicians, which easily leads to a high rate of misdiagnosis and thus delays in treatment. Therefore, exploring a reliable and automated auxiliary diagnostic strategy for autism, especially in its early stages, is receiving increasing attention from the fields of psychiatry and neuroscience.

[0003] Resting-state functional magnetic resonance imaging (fMRI) is a technique that measures blood oxygenation-dependent signals without requiring any specific task to be performed, and it has been widely used for the early diagnosis of autism. Because autism spectrum disorders are found to tend to disrupt neural connections between different brain regions, constructing high-quality functional connectivity networks has become a key issue in capturing subtle abnormal changes in the brain regions of autistic subjects.

[0004] From a mathematical perspective, each node in the brain's functional connectivity network corresponds to a brain region of interest, and each edge corresponds to the interaction between two brain regions. To estimate the edges in the functional connectivity network, researchers have developed many different methods, including Pearson correlation coefficient (PC), sparse representation (SR), mutual information (MI), and correlational correlation (CC).

[0005] These methods can effectively estimate functional connectivity networks, but each can only capture a single type of relationship between different brain regions, making it difficult to model complex interactions in the brain. Inspired by the concept of multi-view learning, fusing complementary features from multiple views has become a popular approach.

[0006] For example, CN115099369A discloses a network fusion method based on functional and structural connectivity. Its technical content includes: dividing the cerebral cortex into different brain regions and networks based on AAL and Yeo brain network templates; preprocessing fMRI and DTI data; constructing a large-scale brain connectivity matrix from the preprocessed fMRI and DTI data; and calculating the network-level weighted probabilities between functional and structural connections. The technical problem is that this proposed network fusion method, based on network-level weighted probabilities, relies solely on the quality of the constructed functional and structural connectivity networks for its classification effect, and fails to consider the differences between different subjects, ignoring the important role of supervisory information in the fusion process.

[0007] For example, Chinese invention patent CN112418337B discloses a multi-feature fusion magnetic resonance imaging data classification method based on a brain functional supernetwork model. It constructs a brain functional supernetwork using the composite MCP method, then uses various indicators to extract features about brain regions to comprehensively quantify the topology of the brain functional supernetwork, and finally applies it to the diagnosis of brain diseases. In this method, when performing multi-feature fusion, the KS test is used to separate fusion and classification based on the distribution of features; subsequent classification does not affect the previous fusion. That is, the quality of the previous fusion does not reflect the fused features and cannot better reflect the true characteristics of the brain. Summary of the Invention

[0008] The technical problem this invention aims to solve is to provide a method for adaptive weighted fusion of multiple functional connectivity networks. This method utilizes supervised information to guide the fusion of multiple functional connectivity networks, automatically learns the weights of each functional connectivity network, and can better solve the problem of fusion weights. The functional connectivity network integrates complementary information from various perspectives, namely information from the perspectives of full correlation, partial correlation, mutual information, and correlational correlation, and integrates brain network fusion and subsequent classification into a single model. The classification results guide the learning of the previous fusion weights. When the learned fusion weights are not good, the learning will continue to be updated. In this way, the fused brain network learned can more realistically reflect the characteristics of the brain.

[0009] To solve the above-mentioned technical problems, the present invention adopts the following technical means:

[0010] A method for adaptive weighted fusion based on multi-functional connectivity networks includes the following steps:

[0011] (1) The brain functional images acquired by the functional magnetic resonance imaging (fMRI) device were preprocessed using the DPARSF toolbox of MATLAB as follows: the first p time points of the magnetic resonance image were removed; head motion and time layer correction were performed on the image; the effects of ventricular, white matter signals and higher-order head motion effects were removed; the corrected image was registered to the standard space; and the image was subjected to a 0.01-0.1 Hz time bandpass filter to reduce the influence of heartbeat and respiration.

[0012] (2) After preprocessing the brain functional image data, the brain was divided into several brain regions based on the automatic anatomical marker atlas, and the average time series of each brain region was extracted and represented as X = [x1, x2, ..., x...]. i , ..., x n ]∈R t×n , where x i ∈R t Let t represent the time series of the i-th brain region, t represent the length of the time series, and n represent the number of brain regions.

[0013] (3) Based on the average time series of brain regions, multiple functional connectivity networks W are created by calculating various correlations between pairs of brain regions. Since the functional connectivity network is mathematically represented as a symmetric matrix, in order to avoid data redundancy, the upper triangular features of the functional connectivity network are selected and flattened to form row vectors.

[0014] (4) Based on the multiple functional connection networks created, each functional connection network adaptively learns the fusion weights using supervised information, thereby obtaining a functional connection network containing multiple correlations.

[0015] (5) The adaptive weighted fusion model of multi-functional connectivity networks is defined as follows:

[0016]

[0017]

[0018] Among them W i j (j = 1, ..., 4) represents the functional connectivity network of the j-th different relevance estimate for the i-th subject, α j W i j The fusion weights, C is the weight vector in L1-norm SVM; y i ∈{-1, 1} is the label corresponding to the i-th subject; M is the number of subjects, K represents the number of functional connectivity networks; λ1 is the error penalty parameter, λ2 is the regularization parameter; ||C||1 tends to make the weight vector C sparse, ||α|| 2 To prevent the model from degenerating to α = (0, 0, 0, 1), constrain α. j ≥0, Not only did it avoid trivial solutions (i.e.) Furthermore, the weight value α is given. j The probabilistic interpretation provides the weight ratio of each relevant functional brain network.

[0019] (6) The model involves two variables, C and α. In this invention, alternating optimization is used to solve the model:

[0020] First: With α fixed, updating C, the model can be simplified to the following objective function L1-norrm SVM:

[0021]

[0022] Many different algorithms have been developed to solve L1-norm SVM. This invention uses the liblinear package in Matlab for implementation.

[0023] Second: With C fixed, and then updating α, the model can be simplified to:

[0024]

[0025]

[0026] Introducing slack variable ξ i If ≥0, the model can be rewritten as:

[0027]

[0028]

[0029] This invention uses the quadratic programming problem solved by combining variables ξ and α. The quadprog function in Matlab is used for the solution.

[0030] The model employs alternating optimization. First, initial weights are given, then the optimal C is optimized. The next step is equivalent to giving C and optimizing the optimal weights, and this process continues alternating. When the difference between the function value of the next iteration and the function value of the previous iteration is less than a certain threshold ε, the algorithm converges, stops iterating, and outputs the optimal function value; otherwise, the calculation is repeated. The weights are the optimal solution obtained by minimizing the model in step (5), i.e., when the difference between the function value of the next iteration and the function value of the previous iteration reaches a certain threshold.

[0031] As a further improvement to this technical solution:

[0032] Step (3) calculates multiple correlations between pairs of brain regions to create multiple functional connectivity networks W, where multiple correlations include full correlation, partial correlation, mutual information and correlation.

[0033] After adopting the above technical solution, the beneficial technical effects of the present invention are: the functional connectivity network in the present invention integrates complementary information from various perspectives, and integrates brain network fusion and subsequent classification into a single model. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating the implementation of the adaptive weighted fusion model based on multi-functional connection networks in this invention. Detailed Implementation

[0035] The present invention will be further described below with reference to the accompanying drawings and specific implementation examples.

[0036] like Figure 1 As shown, a specific implementation example of the adaptive weighted fusion model based on multi-functional connectivity networks includes the following steps:

[0037] (1) In this example, the publicly available Autism Brain Imaging Data Exchange (ABIDE) dataset was used. This dataset contains 184 subjects, including 79 autistic patients and 105 normal individuals. The brain images acquired by the magnetic resonance imaging (MRI) device were read and converted into their format, and preprocessed using the MATLAB DPARSF toolbox. To maintain the stability of the time series, the first 10 time points of each subject's MRI images were removed; then, head movement and time-to-time corrections were applied to the images; the effects of ventricular and white matter signals and higher-order head movement effects were removed; the corrected images were registered to standard space; and a 0.01-0.1 Hz time bandpass filter was applied to the subject images to reduce the influence of heartbeat and respiration.

[0038] (2) Select a standardized brain partitioning template; In this example, the human brain is divided into 116 brain regions (90 brain regions of the cerebrum and 16 brain regions of the cerebellum) according to the AAL template. The 116 brain regions represent 116 nodes of the brain functional connectivity network.

[0039] (3) Calculate the average time series value of all voxels in each brain region in the magnetic resonance image, extract the time series corresponding to different standard partitions, and represent the extracted time series as X = [x1, x2, ..., x...]. i , ..., x n ]∈R t×n , where x i ∈R t Let represent the time series of the i-th node, t represent the length of the time series (i.e., the number of times in the time series), and n represent the number of nodes (i.e., the number of brain regions).

[0040] (4) Based on the average time series of brain regions, multiple functional connectivity networks W are created by calculating various correlations between pairs of brain regions. Since functional connectivity networks are mathematically represented as symmetric matrices, to avoid data redundancy, the upper triangular features of the functional connectivity networks are selected and flattened to form row vectors. Based on the created multiple functional connectivity networks, each functional connectivity network adaptively learns and fuses weights using supervised information, thereby obtaining a functional connectivity network containing multiple correlations.

[0041] (5) The adaptive weighted fusion model of multi-functional connectivity networks is shown below:

[0042]

[0043]

[0044] Among them W i j (j = 1, ..., 4) represents the functional connectivity network of j different correlation estimates for the i-th subject, α j W ij The fusion weights, C is the weight vector in L1-norm SVM; y i ∈{-1, 1} is the label corresponding to the i-th subject; M is the number of subjects, K represents the number of functional connectivity networks; λ1 is the error penalty parameter, λ2 is the regularization parameter; ||C||1 tends to make the weight vector C sparse, ||α|| 2 To prevent the model from degenerating to α = (0, 0, 0, 1), constrain α. j ≥0, Not only did it avoid trivial solutions (i.e.) Furthermore, the weight value α is given. j The probabilistic interpretation.

[0045] (6) The model involves two variables, C and α. In this invention, alternating optimization is used to solve the model:

[0046] First: With α fixed, updating C, the model can be simplified to the following objective function L1-norm SVM:

[0047]

[0048] Many different algorithms have been developed to solve L1-norm SVM. This invention uses the liblinear package in Matlab for implementation.

[0049] Second: With C fixed, and then updating α, the model can be simplified to:

[0050]

[0051]

[0052] Introducing slack variable ξ i If ≥0, the model can be rewritten as:

[0053]

[0054]

[0055] This invention uses the quadratic programming problem solved by combining variables ξ and α. The quadprog function in Matlab is used for the solution.

[0056] (7) When the difference between the function value of the next iteration and the function value of the previous iteration is less than a certain threshold, the threshold is generally chosen to be a sufficiently small number. In this embodiment, ε = 10 is chosen. -6 If the algorithm converges, it stops iterating and outputs the optimal function value; otherwise, it repeats the calculation.

[0057] (8) In step (3), based on the average time series of brain regions, this example uses four types of correlation to calculate the functional connectivity network W between pairs of brain regions, namely full correlation, partial correlation, mutual information and correlation, i.e. K=4 in the weighted fusion model.

[0058] (9) Based on the methods proposed in steps (4)-(7), we integrate a functional connection network with multiple interactive information.

[0059] The effectiveness of the method in this embodiment is verified as follows:

[0060] To verify the effectiveness of the proposed method, we conducted an autism patient vs. normal individuals identification experiment based on an autism brain imaging data exchange dataset. Furthermore, to avoid the influence of complex classifiers on the experimental results, an L1-norm SVM classifier with a default parameter λ1 = 1 was used. In addition, leave-one-out cross-validation was employed to select the optimal parameter λ2. The model effectiveness evaluation metrics included accuracy, sensitivity, specificity, and the area under the receiver operating characteristic curve (AUC). The metrics were defined as follows: Accuracy = (TP + TN) / (TP + FN + FP + TN), Sensitivity = TP / (TP + FN), Specificity = TN / (TN + FP), where TP, TN, FP, and FN represent true positives, true negatives, false positives, and false positives, respectively. Table 1 shows the identification results of six methods for autism spectrum disorder patients under four metrics.

[0061]

[0062] Table 1 shows the classification performance based on five baseline methods (PC, SR, MI, CC, Mean) and the weighted fusion of the multi-functional connectivity network proposed in this invention. This embodiment achieved the best performance across all metrics: Accuracy = 75.44%, Sensitivity = 65.82%, Specificity = 82.86%, and AUC = 79.07%. Furthermore, it was found that MI had poor classification performance; analysis of the fusion weight α revealed that the feature weights provided by MI during the fusion process tended to zero.

[0063] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of the present invention. All equivalent changes made based on the description and drawings of the present invention are included within the scope of the present invention.

Claims

1. A method for adaptive weighted fusion based on multi-functional connectivity networks, characterized in that... Includes the following steps: (1) The brain functional images acquired by the functional magnetic resonance imaging (fMRI) device were preprocessed using the DPARSF toolbox in MATLAB as follows: removing the anterior end of the magnetic resonance image. p At specific time points; head motion and temporal corrections are applied to the images; effects caused by ventricular and white matter signals and higher-order head motion effects are removed; the corrected images are registered to a standard space; and the images are processed using 0.01-0.1... HZ Time bandpass filtering is used to reduce the effects of heartbeat and respiration. (2) After preprocessing the brain functional image data, the brain was divided into several brain regions based on the automatic anatomical marker atlas, and the average time series of each brain region was extracted and represented as follows: ,in Indicates the first The time series of brain regions, among which... Indicates the length of the time series. Indicates the number of brain regions; (3) Based on the average time series of brain regions, multiple functional connectivity networks are created by calculating various correlations between pairs of brain regions. W Since functional connectivity networks are mathematically represented as symmetric matrices, to avoid data redundancy, the upper triangular features of the functional connectivity network are selected and flattened to form row vectors. (4) Based on the multiple functional connection networks created, each functional connection network adaptively learns the fusion weights using supervised information. Thus, a functional connectivity network containing multiple correlations is obtained. ; (5) The adaptive weighted fusion model of multiple related functional connectivity networks is shown below: （1） （2） in, Indicates the first one subject Functional connection networks for different correlation estimates express The fusion weight, C yes L 1 - Weight vectors in a normal SVM; It is the first The labels corresponding to each participant; M For the number of subjects, K Indicates the number of networks connected to the function; It is the error penalty parameter. For regularization parameters; Tendency to make weight vector C Sparse, To prevent the model from degenerating ;constraint ; (6) The model in step (5) involves The model is solved using alternating optimization with two variables: First: Fixed In the case of updating C The model can be simplified to the following objective function. L 1 -norm SVM: (3) Solve using the liblinear package in Matlab. L 1 -norm SVM; Second: Fixed C In the case of, then update The model can be simplified to: （4） （5） Introducing slack variables The above models (4) and (5) can be rewritten as: （6） （7） Using joint variables To solve the quadratic programming problem, use the quadprog function in Matlab. (7) When the difference between the function value of the next iteration and the function value of the previous iteration in step (5) is less than a certain threshold. If the algorithm converges, it stops iterating and outputs the optimal function value; otherwise, it repeats the calculation.

2. The method for adaptive weighted fusion based on multi-functional connectivity networks according to claim 1, characterized in that: Step (3) involves calculating multiple correlations between pairs of brain regions to create multiple functional connectivity networks. W Among them, there are various types of correlations, including full correlation, partial correlation, mutual information, and correlation.

Images