A brain function image classification method based on adaptive high-order fusion learning

Abstract

This invention relates to the field of brain function image classification and processing technology, specifically to an adaptive high-order fusion learning method for brain function image classification, comprising the following steps: S1, acquiring the brain function image to be classified and performing image data preprocessing; S2, constructing a high-order functional brain network matrix sequence; S3, adaptively learning the weights of different-order functional brain networks, weighted summing them to further output features; S4, capturing the contextual dependencies of the high-order functional brain network matrix sequence; S5, further extracting features and performing weighted fusion of features; S6, predicting and outputting the classification result; S7, parameter tuning. This invention, through end-to-end adaptive high-order functional brain network fusion learning, utilizes adaptive learning of different-order functional brain network weights to enhance the interpretability of the inference process between input features and prediction results, and captures temporal information before and after the generation of the high-order functional brain network matrix through a self-attention mechanism, accurately and efficiently classifying and identifying images of normal brain function and brain disorders.

Description

Technical Field

[0001] This invention relates to the field of brain function image classification and recognition technology, specifically to an adaptive high-order fusion learning method for brain function image classification. Background Technology

[0002] Brain diseases are a group of conditions that affect the structure and function of the brain, potentially leading to abnormalities in cognition, emotion, perception, and movement. These diseases can be caused by genetic factors, environmental factors, brain injury, or other physical illnesses. They impose a significant burden on patients' families and society, making research into technologies for early identification and detection crucial.

[0003] Research in neuroscience and clinical medicine has provided important clues for our understanding of the pathophysiological mechanisms of these diseases. Among them, the structured information of the human brain has universal guiding significance for the research and analysis of various brain diseases.

[0004] In recent years, with the rapid development of medical imaging technology, functional magnetic resonance imaging (fMRI) has become a new direction for researchers to study and identify brain diseases. Researchers use brain functional imaging data to construct functional brain networks to characterize each subject. In the functional brain network, each node represents a corresponding brain region, and the edge weights between nodes correspond to the connection strength between two brain regions.

[0005] In existing functional brain network analysis frameworks, researchers typically construct low-order brain networks for learning based on the Pearson Correlation Coefficient (PCC). However, the PCC primarily measures linear relationships and has low sensitivity to nonlinear relationships. This traditional approach to functional brain networks results in insufficient fit to the brain, failing to efficiently represent complex relationships between brain regions and thus affecting the accuracy of downstream classification tasks. Furthermore, while low-order brain network construction methods are effective in identifying certain brain diseases, they are characterized by their inability to capture complex brain interaction patterns. Traditional high-order functional brain networks cannot fully utilize information from the high-level functional brain network construction process and cannot achieve interaction between the feature level and the classification level. Utilizing only a single-order functional brain network undoubtedly leads to information loss, posing challenges to the inference process.

[0006] Therefore, designing and researching a classification method that can accurately identify and classify brain function images to assist in the identification and discovery of brain disorders is an urgent problem to be solved. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides an adaptive high-order fusion learning method for brain functional image classification. It utilizes brain functional images acquired by magnetic resonance imaging (MRI) equipment to construct a high-order functional brain network matrix sequence and adaptively learns the weights of different order functional brain networks. Through a self-attention mechanism, it obtains and fuses multi-order functional brain networks that integrate contextual information, thereby identifying and classifying brain disorder images and normal brain images, which helps in the early identification of brain disorders.

[0008] This invention is achieved through the following technical solution:

[0009] An adaptive high-order fusion learning method for brain function image classification is provided, comprising the following steps:

[0010] S1. Obtain brain function images to be classified and perform image data preprocessing.

[0011] S2. Using the Brain Atlas (AAL), the preprocessed brain functional images are divided into regions of interest (ROIs). The connection weights between these regions are calculated to form a low-order functional brain network. Based on this low-order network, correlations are iteratively calculated to construct a high-order functional brain network matrix sequence. The upper triangular elements of the high-order functional brain network matrix are then flattened to obtain the feature matrix. F .

[0012] S3. Utilizing the constructed high-order functional brain network features, first average the values, then output them to a multilayer perceptron for feature mapping, with the dimension mapped as follows: n Then through The function yields the weight matrix of the higher-order functional brain network. S ; the feature matrix F With weight matrix S Multiplication is used to weight and sum the features of functional brain networks of different orders to obtain an adaptive weighted output. ;

[0013] ;

[0014] in: = [ , , ..., ] , and For a learnable parameter matrix, δ for ReLU Activation function.

[0015] To fully utilize information from functional brain networks of different orders, and addressing the issue of poor interpretability in the inference process between input features and prediction results, an attention-weighted module based on a multilayer perceptron was designed. An adaptive weight learning method was implemented for functional brain networks of different orders, enabling them to focus on features more relevant to the task objective and improving learning ability. By using Pearson correlation calculations to compute the low-order functional brain networks of the subjects, and iteratively performing Pearson correlation calculations, a sequence of high-order functional brain network matrix sequences was constructed. Adaptive learning of the weights for different orders of functional brain networks facilitates the learning of the weight representation best suited to the target task. By adaptively learning weights for the features of functional brain networks of different orders, making them focus more on modules more relevant to the task objective, learning ability is improved, and the interpretability of the inference process between input features and prediction results is enhanced.

[0016] S4. Based on the high-order functional brain network feature matrix in step S2 F A time matrix containing contextual information is obtained through a self-attention mechanism. T , to integrate high-order functional brain network matrix F With time matrix T Multiplication makes the higher-order functional brain network matrix F Learn the contextual information before and after the construction of higher-order functional brain networks;

[0017] ;

[0018] in: d This is represented by the dimension of the extracted features; and For a learnable parameter matrix; according to sum The function sums these features to extract the fused high-order functional brain network features containing temporal information. ;

[0019] .

[0020] Using the high-order functional brain network feature matrix constructed in step S2, a time matrix containing contextual information is obtained through a self-attention mechanism. T By using the feature matrix of high-order functional brain networks F With time matrix T Multiplication captures contextual information of high-order functional brain network matrix sequences before and after generation; summation reveals information from fused functional brain networks of different orders.

[0021] S5, Adjust the adaptive weighted output and fusion output According to hyperparameters α Enter it into the following formula;

[0022]

[0023] S6. Using weighted output O The data is input into a multilayer perceptron, and the predicted classification results are as follows:

[0024] ;

[0025] in: express Function, use The function is normalized to obtain the classification results of brain function image prediction. .

[0026] As an alternative, it also includes S7, parameter tuning, and using brain function images to predict classification results. , To represent the true classification results, hyperparameters are determined using a grid search approach, and the cross-entropy loss function is used.

[0027] ;

[0028] By calculating and controlling the error value of the objective function, the final brain function image classification result is optimized and output.

[0029] This invention is an end-to-end adaptive high-order functional brain network fusion learning process. It utilizes adaptive learning of weights of different order functional brain networks to enhance the interpretability of the inference process between input features and prediction results, improve feature learning ability, enable the method to automatically learn the weight representation most suitable for the target task according to the optimization objective, and capture the temporal information before and after the generation of the high-order functional brain network matrix through a self-attention mechanism to output the final classification result.

[0030] Furthermore, in step S1, image data preprocessing includes removing preprocessing data from brain function images. At each time point, spatial smoothing and temporal bandpass filtering are applied to the images to remove the influence of cerebrospinal fluid and white matter signals, as well as the missing time point data. m The first sample and the first s One sample.

[0031] By preprocessing brain image data, potential instabilities from the initial scan can be eliminated, reducing interference from physiological and instrumental factors, resulting in processed brain functional images and high-quality samples for use.

[0032] Furthermore, in step S2, the low-order functional brain network is described as follows: ,in , Indicates the first N Feature matrix of each brain region Represents a subject, Indicates the number of brain regions. Dimensions representing features , The side feature matrix representing the strength of the linear relationship between two brain regions calculated using Pearson correlation is shown below;

[0033] ;

[0034] in , Indicates the first and the Time series of brain regions yes The average value.

[0035] Furthermore, in step S2, the higher-order functional brain network matrix sequence = [ , , , ], Indicates the first n The functional brain network can be composed of the first-order functional brain network. n The -1 order functional brain network was obtained using Pearson correlation calculations, as shown below;

[0036] ;

[0037] in: C = I For a centered matrix, I It is an identity matrix. E It is a matrix of all ones. diag(A) It is a diagonal matrix, and the matrix A Sharing the same diagonal elements; characteristic matrix F Represented as F = [ , , …, ], Indicates the first n Functional brain network characteristics.

[0038] The beneficial effects of this invention are:

[0039] This invention proposes an adaptive high-order fusion learning-based brain function image classification method. It constructs a sequence of high-order functional brain network matrices based on iterative Pearson correlation calculations and utilizes adaptive learning weights to fully leverage information from different orders of functional brain networks, allowing them to focus more on features more relevant to the target task. Furthermore, to capture the contextual information of the high-order functional brain networks before and after generation, a self-attention mechanism is used to fuse features containing temporal information from these networks. This results in better interpretability and feature learning capabilities, enabling effective recognition of brain disorder images through image classification. The image classification is highly efficient and accurate, contributing to the detection and recognition of brain disorders. Attached Figure Description

[0040] Figure 1 This is a flowchart of the present invention.

[0041] Figure 2 The results are from a comparative experiment using the NYU dataset.

[0042] Figure 3 The results are from a comparative experiment using the MDD dataset.

[0043] Figure 4 The results are presented in this embodiment of the invention, showing the experimental results of using the autism spectrum disorder dataset with four low-order methods, three classical high-order methods, and two deep learning methods.

[0044] Figure 5 The results of the major depressive disorder dataset used in this embodiment of the invention are experimental results using four low-order methods, three classic high-order methods, and two deep learning methods. Detailed Implementation

[0045] To clearly illustrate the technical features of this solution, the following detailed implementation method will be used to explain the solution.

[0046] In this embodiment, resting-state magnetic resonance imaging (MRI) data from two publicly available databases were selected: ABIDE I (Autism Brain Imaging Data Exchange I) and MDD (major depressive disorder). The ABIDE I dataset was initiated by the International Neuroimaging Data Sharing Initiative and comprises data from 17 international imaging centers. The experiment used 840 quality-assessed samples. Within the ABIDE I dataset, NYU, containing the largest sample size (184 samples), was selected for a single-site experiment. The MDD dataset includes neuroimaging data from 280 patients with major depressive disorder and 251 controls with typical developmental stages.

[0047] like Figure 1As shown, an adaptive high-order fusion learning method for brain function image classification includes the following steps:

[0048] S1. Acquire brain functional images to be classified using magnetic resonance imaging (MRI) and perform image data preprocessing.

[0049] Specifically: For brain functional images acquired by functional magnetic resonance imaging (fMRI) before removal At each time point, spatial smoothing and temporal bandpass filtering are applied to the images to eliminate potential instabilities from the initial scan; the influence of cerebrospinal fluid and white matter signals is removed to reduce these physiological and instrumental interferences, resulting in processed brain function images; missing time point data are removed. m The first sample and the first s Only high-quality samples were used; the experiment selected a standard brain atlas (AAL) to divide each subject's brain into sections. N Each region corresponds to a brain region of interest.

[0050] S2. Constructing a high-order functional brain network matrix sequence: For the processed brain functional images obtained in step S1, regions of interest in the brain are divided using the Brain Atlas (AAL), and the connection weights between brain regions are calculated as low-order functional brain networks. The initial low-order functional brain network for each subject is described as follows: ,in , The feature matrix representing the Nth brain region, Represents a subject, Indicates the number of brain regions. Dimensions representing features , The side feature matrix representing the strength of the linear relationship between two brain regions calculated using Pearson correlation is shown below;

[0051] ;

[0052] in , Indicates the first and the Time series of brain regions yes The average value; Pearson correlation calculations are performed iteratively based on the low-order functional brain network of each subject, thereby constructing a sequence of high-order functional brain network matrices. = [ , , , ], This represents the nth-order functional brain network, which can be derived from the nth-order functional brain network. n-1 The functional brain network was obtained using Pearson correlation calculations, as shown below;

[0053] ;

[0054] in: C = I For a centered matrix, I It is an identity matrix. E It is a matrix of all ones. diag(A) It is a diagonal matrix, and the matrix A Sharing the same diagonal elements; flattening the upper triangular elements of the higher-order functional brain network matrix to obtain the feature matrix. F , characteristic matrix F Represented as F = [ , , …, ], Indicates the first n Functional brain network characteristics.

[0055] S3, Adaptive learning of weights for different levels of functional brain networks:

[0056] First, the high-order functional brain network features constructed in step S2 are averaged and output to a multilayer perceptron for feature mapping, mapping its dimensions to... n Then through The function yields the weight matrix of the higher-order functional brain network. S .

[0057] Finally, the feature matrix F With weight matrix S Multiplication involves weighted summation of the features of functional brain networks at different levels of the sample, which is equivalent to weighted aggregation of information from each level of the functional brain network, yielding the output. ;

[0058] ;

[0059] in: = [ , , ..., ] , and For a learnable parameter matrix, δ for ReLU Activation functions; by adaptively learning weights for features of functional brain networks of different orders, they make the network pay more attention to modules that are more relevant to the task objective, thereby improving learning ability and enhancing the interpretability of the inference process between input features and prediction results.

[0060] S4. Capturing the contextual dependencies of high-order functional brain network matrix sequences:

[0061] First, based on the high-order functional brain network feature matrix in step S2, a time matrix containing contextual information is obtained through a self-attention mechanism. T , to integrate high-order functional brain network matrix F With time matrix T Multiplication makes the feature matrix of higher-order functional brain networks F Learn the contextual information before and after the construction of higher-order functional brain networks;

[0062] ;

[0063] in: d This is expressed as the dimension of the extracted features, making its gradient updates more stable during training; and is a learnable parameter matrix.

[0064] Next, according to sum The function sums these features to extract the fused high-order functional brain network features containing temporal information. ;

[0065]

[0066] By computing the self-attention mechanism, we capture the contextual information of the high-order functional brain network matrix sequence before and after its generation, and fuse information from different-order functional brain networks.

[0067] S5. Further extract features and perform weighted fusion: The adaptive weighted output obtained in step S3 is then processed. The fusion output obtained in step S4 According to hyperparameters α Enter it into the following formula;

[0068] .

[0069] S6. Predict the classification result: using the result obtained in step S5. O The input is fed into a multilayer perceptron, and the predicted classification result is as follows:

[0070]

[0071] in: express Function, use The function is normalized to obtain the classification results of brain function image prediction. .

[0072] S7. Parameter tuning: Use the predicted classification results obtained in step S6. , To represent the true classification results, hyperparameters are determined using a grid search approach, and the cross-entropy loss function is used.

[0073]

[0074] By calculating and controlling the error value of the objective function, the final brain function image classification result is optimized and output.

[0075] In this invention, steps S2 to S7 constitute an end-to-end adaptive high-order functional brain network fusion learning process. By adaptively learning the weights of different order functional brain networks, the interpretability of the inference process between input features and prediction results is enhanced, and the feature learning ability is improved. This method enables the method to automatically learn the weight representation most suitable for the target task according to the optimization objective, and captures the temporal information before and after the generation of the high-order functional brain network matrix through a self-attention mechanism, and outputs the final classification result.

[0076] In steps S2 and S3, Pearson correlation is used to calculate the low-order functional brain network of the subject. Pearson correlation is calculated iteratively to construct a sequence of high-order functional brain network matrices. By adaptively learning the weights of different order functional brain networks, it is beneficial to learn the weight representation that is most suitable for the target task.

[0077] In step S4, the high-order functional brain network feature matrix constructed in step S2 is used to obtain a time matrix containing contextual information through a self-attention mechanism. T Then: by using higher-order functional brain network matrices F With time matrix T Multiplication captures contextual information of the high-order functional brain network matrix sequence before and after generation; finally, summation is used to read out the information from different orders of functional brain networks.

[0078] Step S5: The brain function image recognition and classification task is completed by weighted fusion of the adaptive weighted output obtained in step S3 and the fusion output obtained in step S4.

[0079] In the experiments, experiments were conducted on two datasets (ABIDE I, MDD) based on the Brain Atlas (AAL). The experimental results of the classification method of this invention were compared with four classical low-order methods, three classical high-order methods, and two deep learning methods. The method in this embodiment is named "Proposed". The four classical methods are: PC+SVM, SR+SVM, CC2+SVM, and CC3+SVM; the three high-order methods are: GHFBN2, GHFBN3, and HyperGraph; and the two deep learning methods are Transformer and BNT.

[0080] like Figure 4 The results shown are experimental findings using the autism spectrum disorder dataset on four low-order methods, three classic high-order methods, and two deep learning methods.

[0081] like Figure 5 The results shown are experimental findings of using the major depressive disorder dataset on four low-order methods, three classic high-order methods, and two deep learning methods.

[0082] This method is implemented in PyTorch using a NYIDIA GeForce RTX 4060, and the learnable parameters are trained by minimizing the cross-entropy loss function. The final classification layer is configured as a two-layer perceptron. The input features have a dimension of 116×116 on the brain map. The Adam algorithm is used for optimization, with a learning rate of 0.1, adjusted every 5 epochs, a decay factor of 0.5, 100 training iterations, and a batch size of 16.

[0083] In the experiment, the site with the largest sample size, NYU, was selected from the ABIDE I dataset; and the site with the largest sample size, site 20, was selected from the MDD dataset. The experimental results are as follows: Figure 4 and Figure 5 As shown in the analysis, the proposed method achieves high accuracy across several metrics, including Accuracy (ACC), Area Under the Subject Feature Curve (AUC), Precision (PRE), and F1 score, on both datasets. Comparative analysis with four classic low-order methods, three high-order methods, and two deep learning methods demonstrates that this method possesses high classification accuracy.

[0084] In the experiment, such as Figure 4 As shown, experiments were conducted using 184 samples from the ABIDE I dataset. The results demonstrate a significant advantage in classification accuracy compared to four classical low-order methods, three high-order methods, and two deep learning methods. The highest classification accuracy reached 73.34%. Furthermore, considering data imbalance and individual differences among participants, the accuracy and recall of this method were evaluated, yielding an accuracy of 73.24% and a recall of 69.25%. Figure 5 The training results of this method on the MDD dataset are presented. The experimental performance reached 65.36%, and it also performed well on other parameters (such as accuracy, F1 score, and AUC). Figure 4 and Figure 5 The results show that this method achieves high accuracy in various classification tasks and is a method with broad application potential.

[0085] Of course, the above description is not limited to the examples above. Technical features not described in this invention can be implemented by or using existing technology, and will not be repeated here. The above embodiments and drawings are only used to illustrate the technical solutions of this invention and are not intended to limit this invention. This invention has been described in detail with reference to preferred embodiments. Those skilled in the art should understand that any changes, modifications, additions or substitutions made by those skilled in the art within the scope of this invention do not depart from the spirit of this invention and should also fall within the scope of protection of the claims of this invention.

Claims

1. A brain function image classification method based on adaptive high-order fusion learning, characterized in that: Includes the following steps: S1. Obtain brain function images to be classified and perform image data preprocessing; S2. The preprocessed brain functional images are divided into multiple brain regions using the Brain Atlas (AAL). The connection weights between each brain region are calculated as a low-order functional brain network matrix. Based on the low-order functional brain network matrix, the correlation is iteratively calculated to construct a high-order functional brain network matrix sequence. The upper triangular elements of the high-order functional brain network matrix are flattened to obtain the feature matrix. F ; S3. Utilizing the constructed high-order functional brain network features, first average the values, then output them to a multilayer perceptron for feature mapping, with the dimension mapped as follows: k Then through The function obtains the weight matrix of the higher-order functional brain network. S ; the feature matrix F Weight matrix of higher-order functional brain networks S Multiplication is performed to sum the features of functional brain networks of different orders using weighted methods, resulting in an adaptive weighted output. : ； in: = [ , , ..., ] , and For a learnable parameter matrix, δ for ReLU Activation function; S4. Based on the feature matrix of the high-order functional brain network in step S2... F A time matrix containing contextual information is obtained through a self-attention mechanism. Z The feature matrix of high-order functional brain networks F With time matrix Z Multiplication makes the feature matrix of higher-order functional brain networks F Learning from the contextual information before and after the construction of higher-order functional brain networks: ； in: Characteristic matrix transpose, d This is represented by the dimension of the extracted features; and For a learnable parameter matrix; according to sum The function sums these features to extract the fused high-order functional brain network features containing temporal information. : ； S5, Adjust the adaptive weighted output and integration of higher-order functional brain network features According to hyperparameters α Enter it into the following formula: ； S6. Using weighted output The data is input into a multilayer perceptron, and the predicted classification results are as follows: ； in: express Function, use The function is normalized to obtain the classification results of brain function image prediction. .

2. The brain function image classification method based on adaptive high-order fusion learning according to claim 1, characterized in that: It also includes S7, parameter tuning, and using brain function images to predict classification results. ,in To represent the true classification results, hyperparameters are determined using a grid search approach, and the cross-entropy loss function is used. : ； By calculating and controlling the error value of the objective function, the final brain function image classification result is optimized and output.

3. The brain function image classification method based on adaptive high-order fusion learning according to claim 1, characterized in that: In step S1, image data preprocessing includes removing brain function images before processing. At each time point, spatial smoothing and temporal bandpass filtering are applied to the images to remove the influence of cerebrospinal fluid and white matter signals, as well as the missing time point data. m The first sample and the first s One sample.

4. The brain function image classification method based on adaptive high-order fusion learning according to claim 1, characterized in that: In step S2, the low-order functional brain network matrix is ​​described as follows: ,in , Indicates the first N Time-series vectors of brain regions Indicates the sample index. N Indicates the number of brain regions. Dimensions representing features , The edge feature vector representing the strength of the linear relationship between two brain regions calculated using Pearson correlation is shown below: ； in , R M , respectively representing the first The and the first Time series of brain regions yes The average value.

5. The brain function image classification method based on adaptive high-order fusion learning according to claim 4, characterized in that: In step S2, the higher-order functional brain network matrix sequence = [ , , , ], Indicates the first n The functional brain network can be composed of the first-order functional brain network. n The -1 order functional brain network was obtained using Pearson correlation calculations, as shown below: ； in: C = I For a centered matrix, I It is an identity matrix. E It is a matrix of all ones. diag(A) It is a diagonal matrix, and the matrix A Share the same diagonal elements; Feature matrix F Represented as F = [ , , …, ], Indicates the first n Functional brain network characteristics.

Images