A method and system for assisting in the diagnosis of attention deficit hyperactivity disorder

Abstract

The application provides an attention deficit hyperactivity disorder auxiliary diagnosis method and system, and relates to the fields of medical image processing and auxiliary diagnosis. The attention deficit hyperactivity disorder auxiliary diagnosis method comprises the following steps: dividing acquired brain time sequence data into at least one continuous and non-overlapping sub-time window; extracting discriminative features of each sub-time window, and obtaining a high-order functional connection network through the extracted discriminative features; capturing dynamic features of the high-order functional connection network of each sub-time window through a TCN module, and describing the relationship between the functional connection of the brain and time changes; obtaining a value of the correlation between each sub-time window and a disease through the dynamic change rule of the brain disease, and performing feature fusion on the high-order functional connection network of each sub-time window by taking the value of the correlation between each sub-time window and the disease as a weight. The method models the dynamic high-order characteristics of the brain FC, improves the brain functional connection network representation capability, and improves the brain disease auxiliary diagnosis performance.

Description

Technical Field

[0001] This invention relates to the field of medical image processing and auxiliary diagnostic technology, specifically to an auxiliary diagnostic method and system for attention deficit hyperactivity disorder. Background Technology

[0002] Attention deficit hyperactivity disorder (ADHD) is a common neurodevelopmental disorder characterized by inattention, hyperactivity, and impulsivity, which negatively impact a patient's learning, work, and social life. The pathogenesis of ADHD remains unclear, and its diagnosis relies heavily on the patient's medical history and clinical symptoms, making it highly subjective and dependent on the physician's clinical experience. Therefore, identifying reliable and objective diagnostic indicators for early diagnosis and clinical intervention of ADHD is of great significance.

[0003] Resting-state functional magnetic resonance imaging (rs-fMRI), a non-invasive imaging technique, is an important tool for studying brain physiological activity and provides a crucial pathway for uncovering potential disease-related biomarkers in the human brain. Functional connectivity (FC) networks constructed based on rs-fMRI data can characterize neural interactions between brain regions and have been successfully applied to the computer-aided diagnosis of various brain diseases. Studies have shown that abnormal brain network connectivity is present in attention deficit hyperactivity disorder (ADHD), therefore, FC-based methods are widely used in the auxiliary diagnosis of ADHD.

[0004] Traditional full-connection (FC) network analysis typically assumes that rs-fMRI is stable throughout the scan. However, recent studies have shown that functional brain connectivity exhibits dynamic changes even at rest. Revealing these dynamic changes can help identify more sensitive imaging biomarkers, providing a new perspective and approach for understanding brain diseases. Existing research often constructs dynamic FCs based on the sliding window technique. This involves dividing the entire time series into multiple non-overlapping sub-time windows, calculating the statistical dependencies of time-series signals between paired brain regions within each window to construct a functional connectivity network, and analyzing its dynamic changes over time. However, previous studies have only focused on the learning of local relationships between paired brain regions, neglecting higher-order interactions in spontaneous brain activity. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] To address the shortcomings of existing technologies, this invention provides an auxiliary diagnostic method and system for attention deficit hyperactivity disorder (ADHD), which models the dynamic higher-order characteristics of the brain's functional connectivity network (FC), improves the representation ability of brain functional connectivity networks, and enhances the auxiliary diagnostic performance of brain diseases.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] Firstly, a method for assisting in the diagnosis of attention deficit hyperactivity disorder is provided, including:

[0010] The acquired brain time-series data is divided into at least one continuous and non-overlapping sub-time window;

[0011] Discriminant features are extracted from each sub-time window, and high-order functional connectivity networks are obtained through the extracted discriminant features;

[0012] By capturing the dynamic characteristics of the high-order functional connectivity network of each sub-time window through the preset TCN module, the relationship between brain functional connectivity and changes over time can be characterized, and the dynamic change pattern of brain diseases can be obtained.

[0013] By obtaining the correlation values ​​between each sub-time window and the disease through the dynamic changes of brain diseases, and using the correlation values ​​between each sub-time window and the disease as weights, feature fusion is performed on the high-order functional connectivity network of each sub-time window;

[0014] The result of feature fusion is input into a three-layer fully connected layer for classification transformation to obtain the brain disease prediction classification result.

[0015] Preferably, the acquired brain time-series data is divided into at least one continuous and non-overlapping sub-time windows:

[0016] We acquired functional magnetic resonance brain imaging (fMRI) data of subjects at rest, preprocessed the fMRI data, and obtained time series of blood oxygenation level-dependent signals in each region of interest.

[0017] Using the sliding window technique, time series data is divided into T consecutive and non-overlapping sub-time windows, each with a length of L.

[0018] Preferably, the step of extracting discriminative features for each sub-time window and obtaining a higher-order functional connectivity network through the extracted discriminative features specifically includes:

[0019] For the t-th sub-window, a more discriminative data representation is obtained using a three-layer CNN convolution operation. The specific calculation formula is as follows:

[0020]

[0021] in, This is the k-th feature map of brain region i, where * denotes a convolution operation, W k This represents the weight of the k-th filter, with a width of L and b. k It is a bias term; is the time-series data of brain region i within the t-th time window; σ represents the activation function; the CNN convolutional module has three layers, each performing BatchNormalization, ReLU activation, and Dropout operations, with each layer having f1, f2, and f3 filters, and each filter having lengths d1, d2, and d3, respectively. Then, bilinear pooling is performed to obtain the higher-order functional connectivity network, the specific calculation formula of which is as follows:

[0022]

[0023] Among them, F t This represents the output of the CNN convolution in the t-th window. H is the tensor representation of brain region i within the t-th window; H is the constructed high-order functional connectivity network, which is the overall representation of the feature space and is used to capture second-order statistical information of feature activation to achieve a representation of the brain.

[0024] Preferably, the dynamic features of the high-order functional connectivity network captured by the preset TCN module for each sub-time window are used to characterize the relationship between brain functional connectivity changes over time and obtain the dynamic change pattern of brain diseases. Specifically, this includes:

[0025] Extracting the high-order function connectivity network H for each sub-time window t The upper triangular element connects the higher-order functions of each sub-time window to the network H. t The upper triangular elements are stretched into vector form and then used to capture the dynamic characteristics of brain functional connectivity networks through the TCN module. In the TCN module, each layer employs dilated causal convolution, followed by Batch Normalization, ReLU activation, and Dropout operations, and further uses residual connections to facilitate gradient flow. The calculation formula for each layer is as follows:

[0026]

[0027]

[0028] in, It is the output of the t-th window at the l-th layer. * indicates a convolution operation. W1 and W2 are dilation factor convolutions with a kernel size of 3 and a number of convolution filters of f4.

[0029] Extract the generated high-order functional connection network N×T×R×R four-dimensional tensor data, flatten the last two dimensions of the four-dimensional tensor data into vectors, and the four-dimensional tensor becomes N×T×(R×(R-1) / 2) three-dimensional tensor data. After being sent into the TCN module, the shape of the three-dimensional tensor data becomes N×T×f4, where f4 is the number of convolutional filters.

[0030] Preferably, the step of obtaining the correlation value between each sub-time window and the disease through the dynamic changes of brain diseases, and using the correlation value between each sub-time window and the disease as a weight to perform feature fusion on the higher-order functional connectivity network of each sub-time window, specifically includes:

[0031] Applying a 1D 1×1 convolution to the output of the TCN module to obtain the correlation between each sub-time window and the disease, multiplying the obtained values ​​with the input of the TCN module and fusing them, we obtain dynamic high-order brain network fusion features.

[0032] By using 1D 1×1 convolutions, the expressive power of the model can be further enhanced without increasing computational cost.

[0033] N×T×f4 tensor data is transformed into N×T×1 tensor data after 1D 1×1 convolution. After further channel compression, N×T 2D tensor data is obtained. The N×T 2D tensor data represents the correlation between each window and the disease.

[0034] Preferably, the step of inputting the feature fusion result into a three-layer fully connected layer for classification transformation to obtain the brain disease prediction classification result specifically includes:

[0035] The sizes of the three fully connected layers are 512, 128 and 2 respectively. The result of feature fusion is fed into the Softmax function after passing through the three fully connected layers to realize the transformation of the classification output value.

[0036] The N×T tensor data is transformed into N×512, N×128 and N×2 respectively through three fully connected layers; the N×2 is fed into the Softmax function, so that the elements in the two-dimensional tensor are mapped to the binary space {0,1}.

[0037] Secondly, an auxiliary diagnostic system for attention deficit hyperactivity disorder is provided, the system comprising the following modules:

[0038] The segmentation module is used to divide the acquired brain time-series data into at least one continuous and non-overlapping sub-time windows;

[0039] The first processing module is used to extract the discriminative features of each sub-time window and obtain the higher-order functional connection network through the extracted discriminative features;

[0040] The second processing module is used to capture the dynamic features of the high-order functional connectivity network of each sub-time window obtained through the preset TCN module, characterize the relationship of brain functional connectivity changes over time, and obtain the dynamic change pattern of brain diseases.

[0041] The feature fusion module is used to obtain the correlation value between each sub-time window and the disease through the dynamic changes of brain diseases, and to use the correlation value between each sub-time window and the disease as a weight to perform feature fusion on the high-order functional connectivity network of each sub-time window.

[0042] The classification module is used to input the result of feature fusion into a three-layer fully connected layer for classification transformation, so as to obtain the brain disease prediction classification result.

[0043] Thirdly, a computer-readable storage medium is provided for storing one or more programs, the one or more programs including instructions that, when executed by a computing device, cause the computing device to perform any of the methods described.

[0044] Fourthly, a computing device is provided, comprising:

[0045] One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include instructions for performing any of the methods described.

[0046] (III) Beneficial Effects

[0047] This invention provides an auxiliary diagnostic method and system for attention deficit hyperactivity disorder (ADHD). Unlike static functional connectivity analysis methods, it constructs a dynamic higher-order functional connectivity brain network based on a sliding window approach. Unlike traditional algorithms that treat feature extraction and classifier learning as two separate learning tasks, this method unifies feature extraction and classifier learning into an end-to-end deep model. It designs CNN and TCN modules, fusing the dynamic higher-order characteristics of functional connectivity for brain disease classification. This invention can assist doctors in diagnosing ADHD, facilitate the quantitative assessment of ADHD, and provide doctors with a means of clinical auxiliary diagnosis. Attached Figure Description

[0048] Figure 1 This is a flowchart of the method of the present invention.

[0049] Figure 2 This is a graph showing the inter-group differences in the FC learned by the KKI site in this embodiment of the invention.

[0050] Figure 3 This is a diagram showing the dynamic change pattern learned by the HT-Net method in this embodiment of the invention. Detailed Implementation

[0051] The technical solutions in the embodiments of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. 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.

[0052] Example

[0053] like Figure 1 As shown, this embodiment of the invention provides an auxiliary diagnostic method for attention deficit hyperactivity disorder, including:

[0054] Step (1): Divide the acquired brain time series data into at least one continuous and non-overlapping sub-time window;

[0055] Step (2): Extract the discriminative features of each sub-time window, and obtain the high-order functional connectivity network through the extracted discriminative features;

[0056] Step (3): Capture the dynamic features of the high-order functional connectivity network of each sub-time window through the preset TCN module, characterize the relationship of brain functional connectivity changes over time, and obtain the dynamic change pattern of brain diseases.

[0057] Step (4): Obtain the correlation value between each sub-time window and the disease through the dynamic change pattern of brain diseases, and use the correlation value between each sub-time window and the disease as weight to perform feature fusion on the high-order functional connectivity network of each sub-time window;

[0058] Step (5): Input the result of feature fusion into a three-layer fully connected layer for classification conversion to obtain the brain disease prediction classification result.

[0059] Furthermore, in step (1), for each sample, the brain time series data is divided into T consecutive and non-overlapping sub-time windows using the sliding window technique;

[0060] We acquired functional magnetic resonance brain imaging data of subjects at rest, and then preprocessed them to obtain time series of blood oxygenation level-dependent signals in each region of interest.

[0061] To ensure the stability of the signal, the first 4 time points are removed for magnetization equilibrium. The preprocessing of functional magnetic resonance brain imaging data mainly includes the following steps: (1) Temporal correction. Since the scanning order is slice-by-slice scanning, a temporal correction is performed for each slice during acquisition; (2) Head motion correction. Each frame image in the experimental sequence is aligned with the middle frame image of this sequence according to a certain algorithm to correct head motion; (3) Spatial normalization. The images acquired from different individuals are uniformly registered onto a standard brain atlas (Montreal Neurological Institute, MNI) template for subsequent processing. (4) Spatial smoothing. To reduce the noise impact on the images and blur the data, thereby improving the signal-to-noise ratio, the images are spatially smoothed with a Gaussian kernel of 6 mm; (5) Spatial filtering and temporal signal extraction. The signal of functional magnetic resonance belongs to low-frequency signals, so the signal is band-pass filtered (0.009 Hz < f < 0.08 Hz), and the temporal series signal is extracted according to brain regions.

[0062] Then, the sliding window technique is used, and the time series data is divided into T sub-time windows of equal length, continuous, and non-overlapping using a sub-sliding window of length L.

[0063] In each embodiment, the data becomes tensor data of T groups of N×R×L shapes through the sliding window technique. Here, N is the number of samples, and R represents the number of regions of interest.

[0064] Furthermore, in step (2), discriminative representations of each sliding window time series are extracted through three-layer CNN convolution, and a high-order functional connection network is constructed using bilinear pooling operation;

[0065] To be able to use multi-channel convolution operations, the original data is dimensionally elevated and reshaped into tensor data of T groups of N×C×R×L shapes. The data is fed into three-layer convolution operations to obtain a more discriminative data representation. The calculation formula for the three-layer convolution is as follows:

[0066]

[0067] Among them, is the k-th feature map of brain region i, * represents the convolution operation, W k represents the weight of the k-th filter, whose width is L, and b k is the bias term. is the time series data of brain region i within the t-th time window. σ represents the activation function. The CNN convolution module has three layers, and each layer performs convolution operation, Batch Normalization, ReLU activation, and Dropout operation respectively. The number of filters in each layer is f1, f2, and f3 respectively, and the length of each filter is d1, d2, and 1 respectively.

[0068] In each embodiment, after three layers of CNN convolution, the data is transformed into T sets of tensor data of shape N×1×R×L', where L' represents the length of the sliding window after convolution.

[0069] Then, bilinear pooling is performed to obtain the higher-order functional connectivity network, the specific calculation formula of which is as follows:

[0070]

[0071] Among them, F t This represents the output of the CNN convolution in the t-th window. is the tensor representation of brain region i within the t-th window. H is the constructed high-order functional connectivity network, which is the overall representation of the feature space and can capture second-order statistical information of feature activation, thus enabling a better representation of the brain.

[0072] In each embodiment, after bilinear pooling, T sets of tensor data of shape N×1×R×R are obtained. The Concatenate function is used to concatenate the T sets of tensor data together according to the second dimension to generate a higher-dimensional tensor, that is, to generate an N×T×R×R high-order function connection network.

[0073] Furthermore, step (3) uses the TCN module to characterize the relationship between brain functional connectivity changes over time in order to obtain the dynamic change pattern of brain diseases;

[0074] The upper triangular elements of the high-order functional connectivity network (Ht) constructed by each sliding window are extracted, stretched into vector form, and then the dynamic characteristics of the brain's functional connectivity network are captured through the TCN module. Unlike traditional CNN convolutions, each layer in the TCN module uses dilated causal convolution, followed by Batch Normalization, ReLU activation, and Dropout operations, and further uses residual connections to facilitate gradient flow. The calculation formula for each layer is as follows:

[0075]

[0076]

[0077] in, This is the output of the t-th window at layer l. * indicates a convolution operation. W1 and W2 are dilation factor convolutions with a kernel size of K. TThe number of convolutional filters is f4. Using dilation factor convolution increases the receptive field without increasing the number of parameters by increasing the number of layers or kernel size. Three layers sequentially stack multiple TCN modules, allowing for progressive improvement of predictions from the earlier stages. Since the output of each stage is an initial prediction, the network is able to capture dependencies between functional connections at different time points and learn reasonable dynamic changes.

[0078] In each embodiment, the generated high-order functional connection network N×T×R×R four-dimensional tensor data is extracted, and the last two dimensions are flattened into vectors. At this time, the four-dimensional tensor becomes N×T×(R×(R-1) / 2) three-dimensional tensor data. After being sent into the TCN module, the shape of the three-dimensional tensor data becomes N×T×f4, where f4 is the number of convolutional filters.

[0079] Furthermore, in step (4), a 1D 1×1 convolution is applied to the output tensor of the TCN module to obtain the correlation between each window and the disease, and this correlation is used as a weight to perform feature fusion on the functional connection network of multiple sliding windows.

[0080] A 1D 1×1 convolution is applied to the output of the TCN module to obtain the correlation between each sub-window and the disease. The obtained values ​​are multiplied with the input of the TCN module and fused to obtain dynamic high-order brain network fusion features.

[0081] By using 1D 1×1 convolutions, the expressive power of the model can be further enhanced without increasing computational cost.

[0082] In each embodiment, N×T×f4 tensor data is transformed into N×T×1 tensor data after 1D 1×1 convolution. After further channel compression, N×T two-dimensional tensor data is obtained. This two-dimensional tensor data represents the correlation between each window and the disease. For example, Ni×Tj represents the correlation between the j-th time window in the i-th sample and the diagnosis of attention deficit hyperactivity disorder.

[0083] Furthermore, in step (5), the dynamic high-order brain network fusion features are input into a three-layer fully connected layer, and Softmax is used as the activation function to obtain the brain disease prediction and classification results.

[0084] The three fully connected layers have sizes of 512, 128, and 2, respectively. After passing through the fully connected layers, the data is fed into the Softmax function to convert the classification output values.

[0085] In each embodiment, the N×T tensor data is transformed into N×512, N×128, and N×2 through three fully connected layers, respectively. Finally, the N×2 tensor is fed into the Softmax function, which maps the elements of the two-dimensional tensor to the binary space of {0, 1}.

[0086] In a specific embodiment, using the ADHD-200 dataset as training samples, the steps of the dynamic higher-order brain network for auxiliary diagnostic classification in the dataset are illustrated, including the following steps:

[0087] Step 1: Obtain the data.

[0088] Download and prepare the ADHD-200 dataset.

[0089] We conducted our experiments on the publicly available ADHD-200 dataset, which comprises data from eight sites: the Kennedy Krieger Institute (KKI), NeuroImage, New York University Child Study Center (NYU), Oregon Health & Science University (OHSU), Peking University Child Study Center (PKU), University of Pittsburgh (UPittsburgh), University of Washington (WashU), and Brown University (BU). The BU site data did not provide participant labels, while the UPittsburgh and WashU sites only provided control group data in their training datasets. Therefore, this paper uses data from the first five sites for our experiments. Table 1 presents the clinical information of the ADHD-200 data.

[0090] Step 2: Use the sliding window technique to segment brain time-series data.

[0091] The acquired data has been divided into training and testing sets, and at this stage, the data is a tensor of shape N×R×S. Then, the brain time-series data is divided into T consecutive and non-overlapping sub-time windows using the sliding window technique, resulting in T groups of tensors of shape N×R×L. For subsequent multi-channel convolution operations, the data is upscaled to obtain T groups of tensor data of shape N×R×L×1.

[0092] Step 3: Establish an auxiliary diagnostic classification network.

[0093] The network mainly consists of CNN modules, bilinear pooling, TCN modules, and feature fusion. The CNN convolutional module has three layers, each performing convolution, batch normalization, ReLU activation, and dropout operations. Bilinear pooling is used to construct a high-order functional connectivity network, capturing second-order statistical information from feature activations for better brain representation. Each layer of the TCN module uses dilated causal convolution, followed by batch normalization, ReLU activation, and dropout operations, further utilizing residual connections to facilitate gradient flow. The feature fusion operation applies a 1D 1×1 convolution to the TCN module output to obtain the correlation between each sub-window and the disease. This value is then multiplied by the TCN module input and fused to obtain dynamic high-order brain network fusion features. Finally, the resulting data is fed into three fully connected layers, using Softmax to transform the classification output values.

[0094] Step 4: Set network parameters.

[0095] The network was trained in Tensorflow 2.6.2 and Keras 2.6.0 environments with a training batch size of 16 and dropout of 0.3, resulting in a trained high-order brain network for auxiliary diagnosis and classification.

[0096] Step 5: Use the trained high-order brain network to perform classification and prediction of the samples.

[0097] Through the learning and training in steps 3 and 4, the optimal network model parameters are obtained by manually tuning the parameters, and the class prediction is performed on the test sample data in the dataset.

[0098] We tested the proposed method on five sites (KKI, NeuroIMAGE, NYU, OHSU, and PKU) of the ADHD-200 public dataset. This invention performs an ADHD vs. NC classification task. Classification accuracy is used as the metric to evaluate the performance of different methods. To verify the effectiveness of the proposed model (HT-Net), baseline methods based on static fully connected (FC) networks, namely MLP and CNN models, were also compared. Furthermore, methods utilizing the dynamic characteristics of brain functional connectivity, namely LSTM and CNN-LSTM models, were also compared. The comparison results of the five methods are shown in Table 2.

[0099] Table 1. Statistical information of the ADHD-200 dataset sample

[0100]

[0101] Table 2. Comparison results of five methods in the ADHD vs. NC classification task.

[0102]

[0103] As shown in Table 2, the proposed method exhibits better classification performance compared to the other four methods. Specifically, the proposed method outperforms models based on statically functionally connected networks (i.e., MLP and CNN), indicating that higher-order fully connected (FC) networks can effectively improve diagnostic performance. Furthermore, the proposed method outperforms models based on dynamically functionally connected networks (i.e., LSTM and CNN-LSTM), demonstrating the advantage of integrating higher-order information and dynamic characteristics. To verify the superior representational power of the learned higher-order FC networks, inter-group differences in the learned FC networks at KKI sites are illustrated, such as... Figure 2 As shown, there are five time points in the high-order fully connected (FC) network. We performed inter-group comparisons on all samples at each time point, with darker colors indicating significant differences between groups (i.e., p < 0.05). The results show that the learned high-order FC network has better discriminative power. Simultaneously, the dynamic change patterns learned by the HT-Net method are demonstrated, such as... Figure 3 As shown in the results, all test samples had similar importance at time points 1, 4 and 5, while there were some differences at time points 2 and 3, which may be closely related to abnormalities in the development of brain diseases.

[0104] Embodiments of this application may be provided as methods or computer program products. Therefore, this application may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application may be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0105] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0106] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0107] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0108] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A method for assisting in the diagnosis of attention deficit hyperactivity disorder, characterized in that, include: The acquired brain time-series data is divided into at least one continuous and non-overlapping sub-time window; Discriminant features are extracted from each sub-time window, and high-order functional connectivity networks are obtained through the extracted discriminant features; By capturing the dynamic characteristics of the higher-order functional connectivity network in each sub-time window through the preset TCN module, the relationship between brain functional connectivity and changes over time can be characterized, and the dynamic change pattern of brain diseases can be obtained. By obtaining the correlation values ​​between each sub-time window and the disease through the dynamic changes of brain diseases, and using the correlation values ​​between each sub-time window and the disease as weights, feature fusion is performed on the high-order functional connectivity network of each sub-time window; The result of feature fusion is input into a three-layer fully connected layer for classification transformation to obtain the brain disease prediction classification result; The step of extracting discriminative features for each sub-time window and obtaining a high-order functional connectivity network using the extracted discriminative features specifically includes: For the t-th sub-window, a more discriminative data representation is obtained using a three-layer CNN convolution operation. The specific calculation formula is as follows: in, It is the k-th feature map of brain region i. This represents the convolution operation. This represents the weight of the k-th filter, with a width of L. It is a bias term; This is the time-series data of brain region i within the t-th time window; This represents the activation function; the CNN convolutional module has three layers, each performing Batch Normalization, ReLU activation, and Dropout operations, with the number of filters in each layer being... , and The length of each filter is respectively , and Then, bilinear pooling is performed to obtain the higher-order functional connectivity network, the specific calculation formula of which is as follows: in, This represents the output of the CNN convolution in the t-th window. H is the tensor representation of brain region i within the t-th window; H is the constructed high-order functional connectivity network, which is the overall representation of the feature space and is used to capture second-order statistical information of feature activation to achieve a representation of the brain. The dynamic characteristics of the high-order functional connectivity network captured by the preset TCN module for each sub-time window are used to characterize the relationship of brain functional connectivity changes over time, thereby obtaining the dynamic change patterns of brain diseases. Specifically, this includes: Extracting the high-order function connectivity network for each sub-time window The upper triangular element connects the higher-order functions of each sub-time window to the network. The upper triangular elements are stretched into vector form and then used to capture the dynamic characteristics of brain functional connectivity networks through the TCN module. In the TCN module, each layer employs dilated causal convolution, followed by Batch Normalization, ReLU activation, and Dropout operations, and further utilizes residual connections to facilitate gradient flow. The calculation formula for each layer is as follows: in, It is the output of the t-th window at the l-th layer. This represents the convolution operation. and It is a dilation factor convolution with a kernel size of 3 and a number of convolution filters. indivual; The generated high-order functional connectivity network (HFM) of N×T×R×R four-dimensional tensor data is extracted. The last two dimensions of the four-dimensional tensor data are flattened into vectors, transforming the four-dimensional tensor into N×T×(R×(R-1) / 2) three-dimensional tensor data. After being fed into the TCN module, the shape of the three-dimensional tensor data becomes N×T× ,in It represents the number of convolutional filters; The process involves obtaining the correlation values ​​between each sub-time window and the disease based on the dynamic changes in brain diseases, and then using these correlation values ​​as weights to perform feature fusion on the high-order functional connectivity network of each sub-time window. Specifically, this includes: Applying a 1D 1×1 convolution to the output of the TCN module to obtain the correlation between each sub-time window and the disease, multiplying the obtained values ​​with the input of the TCN module and fusing them, we obtain dynamic high-order brain network fusion features. By using 1D 1×1 convolution, the expressive power of the model is further enhanced without increasing the computational cost; N×T×f4 tensor data is transformed into tensor data of shape N×T×1 after 1D 1×1 convolution, and after further channel compression, N×T shaped two-dimensional tensor data is obtained. The N×T shaped two-dimensional tensor data represents the correlation between each window and the disease. The result of feature fusion is input into a three-layer fully connected layer for classification transformation to obtain the brain disease prediction classification result.

2. The auxiliary diagnostic method for attention deficit hyperactivity disorder according to claim 1, characterized in that: The acquired brain time-series data is divided into at least one continuous and non-overlapping sub-time window: We acquired functional magnetic resonance brain imaging (fMRI) data of subjects at rest, preprocessed the fMRI data, and obtained time series of blood oxygenation level-dependent signals in each region of interest. Using the sliding window technique, time series data is divided into T consecutive and non-overlapping sub-time windows, each with a length of L.

3. The auxiliary diagnostic method for attention deficit hyperactivity disorder according to claim 1, characterized in that: The process of inputting the feature fusion result into a three-layer fully connected layer for classification transformation to obtain the brain disease prediction classification result specifically includes: The sizes of the three fully connected layers are 512, 128 and 2 respectively. The result of feature fusion is fed into the Softmax function after passing through the three fully connected layers to realize the transformation of the classification output value. The N×T tensor data is transformed into N×512, N×128 and N×2 respectively through three fully connected layers; the N×2 is fed into the Softmax function, so that the elements in the two-dimensional tensor are mapped to the binary space {0,1}.

4. A system for performing the auxiliary diagnostic method for attention deficit hyperactivity disorder according to any one of claims 1-3, characterized in that, The system includes the following modules: The segmentation module is used to divide the acquired brain time-series data into at least one continuous and non-overlapping sub-time windows; The first processing module is used to extract the discriminative features of each sub-time window and obtain the higher-order functional connection network through the extracted discriminative features; The second processing module is used to capture the dynamic characteristics of the higher-order functional connectivity network of each sub-time window through the preset TCN module, characterize the relationship of brain functional connectivity changes over time, and obtain the dynamic change pattern of brain diseases. The feature fusion module is used to obtain the correlation value between each sub-time window and the disease through the dynamic changes of brain diseases, and to use the correlation value between each sub-time window and the disease as a weight to perform feature fusion on the high-order functional connectivity network of each sub-time window. The classification module is used to input the result of feature fusion into a three-layer fully connected layer for classification transformation, so as to obtain the brain disease prediction classification result.

5. A computer-readable storage medium for storing one or more programs, characterized in that, The one or more programs include instructions that, when executed by a computing device, cause the computing device to perform any of the methods according to claims 1-3.

6. A computing device, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs including instructions for performing any of the methods according to claims 1-3.

Images