Attention deficit hyperactivity disorder auxiliary diagnosis method based on field generalization

By adopting a domain generalization method in ADHD diagnosis, the common and unique features of multiple medical center data sets are extracted and the attention network is used to ensure the interpretability of features, which solves the problem of reduced diagnostic accuracy caused by data heterogeneity and improves the The generalization ability and medical explanatory nature of the model are achieved, and the diagnosis results with high accuracy are achieved.

CN119028571APending Publication Date: 2024-11-26HOHAI UNIV
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411124184.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-15
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

When using deep learning models to diagnose attention deficit hyperactivity disorder (ADHD), the prior art is affected by data heterogeneity, resulting in a decrease in diagnostic accuracy, and the characteristics extracted by the model lack medical interpretability, which affects clinical applications.

Method used

Using a domain generalization-based method, we extract the common and unique features of multiple medical center data sets, and design the attention network through the attention mechanism to ensure the interpretability of feature maps, thereby enhancing the generalization ability and medical ability of the model. Explanatory.

Benefits of technology

Improve the diagnostic accuracy and generalization ability of the model, identify key significant brain region features, enhance the reliability and medical interpretability of the diagnosis, and achieve a classification accuracy of 72.8% on the ADHD-200 dataset. .

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119028571A_ABST
    Figure CN119028571A_ABST
Patent Text Reader

Abstract

The invention provides an attention deficit hyperactivity disorder auxiliary diagnosis method based on domain generalization. An auxiliary diagnosis model is formed by deep learning modules such as an attention network, a classification network and a coding and decoding network. According to the method, common features and specific features of data from multiple medical centers are considered, the common features of multiple data sets are extracted to align different data distributions, and meanwhile, the specific features of the data sets are extracted to enhance the generalization ability; and setting an attention module to ensure the interpretability of the mapping features, extracting the attention deficit hyperactivity disorder biomarker according to the interpretability of the mapping features, and enhancing the medical interpretability of the model. A test result on an ADHD-200 data set shows that 72.8% of classification accuracy is obtained on a new data set with unknown training under the condition of more training data, and the obtained significant brain region conforms to a medical result of an existing attention deficit hyperactivity disorder disease.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of intelligent auxiliary diagnosis of mental illness and domain generalization technology, and in particular to an auxiliary diagnosis method for attention deficit hyperactivity disorder based on domain generalization. Background Art

[0002] Attention Deficit Hyperactivity Disorder (ADHD), commonly known as ADHD, is a common neurodevelopmental disorder in children. Currently, clinical diagnosis is still mainly based on subjective diagnosis, with common methods being clinically based structured interviews and behavioral scale scores. There is a lack of objective diagnostic methods to accurately evaluate brain abnormalities in children, and it is also impossible to conduct a more in-depth study of the brain pathogenesis of ADHD.

[0003] With the development of artificial intelligence algorithms such as deep learning, intelligent auxiliary diagnosis methods using individual neurobiological representation data have become a research hotspot. Among them, due to the strong learning and representation ability of deep neural networks, they can improve the learning ability of complex nonlinear ADHD brain imaging data features, thus achieving higher classification accuracy than traditional machine learning methods.

[0004] Through patent search, it is found that the number of patents related to ADHD intelligent assisted diagnosis and treatment is relatively small. The specific patent information is listed as follows [1]-[4]:

[0005] [1] Li Wenjie, Han Qiulei, Zhao Jian, Kuang Zhejun. An auxiliary diagnosis system for attention deficit hyperactivity disorder syndrome based on a comprehensive cognitive experiment set and a multi-stream double-row decision neural network model [P]. Jilin Province: CN117357114A, 2024-01-09.

[0006] [2] Jiang Xinlong, Huang Wuliang, Chen Yiqiang, Zhang Teng, Xing Yunbing. An auxiliary diagnostic device for attention deficit hyperactivity disorder and its use method [P]. Beijing: CN110693510A, 2020-01-17.

[0007] [3] Wang Mingliang, Zhu Lingyao, Li Xizhi. A method and system for auxiliary diagnosis of attention deficit hyperactivity disorder [P]. Jiangsu Province: CN116889405A, 2023-10-17.

[0008] [4] Jiang Xinlong, Chen Yiqiang, Zheng Yi, Qiao Yuxi, Huang Wuliang, Zhou Yuming. Method and system for identifying subtypes of attention deficit hyperactivity disorder [P]. Beijing: CN113160967A, 2021-07-23.

[0009] However, existing patents have not considered that the imaging equipment used to collect ADHD brain imaging data may come from different manufacturers, and there may be differences in equipment parameter settings, equipment models and performance, resulting in differences in the distribution of data collected by different medical centers, namely data heterogeneity. This will affect the generalization ability of the learning model, that is, the model trained with data from one or some medical centers will show a significantly lower diagnostic accuracy for input data from other medical centers.

[0010] At present, the common methods to solve the problem of data heterogeneity affecting the generalization ability of deep learning models include domain adaptation and domain generalization. Domain adaptation is to migrate the features of a known data set (source domain) to the target domain (i.e., target center data). It requires the target domain to be known and participate in model training. Therefore, the migrated features are only applicable to the known target domain and are difficult to adapt to new unknown target domains. Domain generalization is to train the model on multiple source domains and make the model applicable to new unknown target domains without retraining. For ADHD disease diagnosis, the unknown target domain means that the brain imaging data of the diagnosed person is unknown during the diagnostic model training stage, which is in line with the actual medical diagnosis scenario. Therefore, domain generalization is more suitable for ADHD disease diagnosis deep learning models with multi-center data.

[0011] There are successful precedents for applying neighborhood generalization to intelligent diagnosis of mental illness. For example, the literature [5] (Lee J, Kang E, Jeon E, et al. Meta-modulation Network for Domain Generalization in Multi-site fMRI Classification [C], International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, Cham, 2021: 500-509.) uses a domain generalization model proposed by a meta-learning strategy to diagnose autism based on multi-center autism brain imaging data. The model achieves a classification accuracy similar to that of the source domain in the target domain. For ADHD, the domain generalization method has been applied to the study of ADHD diagnosis based on multi-center data of EEG signals, and the classification accuracy in the target domain is similar to or slightly higher than that of the source domain [6] (Qin X, Wang J, Chen Y, et al. Domain Generalization for Activity Recognition via Adaptive Feature Fusion [J]. ACM Transactions on Intelligent Systems and Technology, 2022, 14 (1): 1-21). However, these domain generalization methods have no restrictions on the mapping space of features and cannot guarantee the medical interpretability of the mapped features, which is not conducive to the identification of ADHD biomarkers and thus affects the actual clinical application of the model.

[0012] Domain generalization methods usually learn domain-invariant features of known source domains. In fact, multiple source domains and target domains have their own unique features. Unique features are equivalent to the unique attributes that distinguish a dataset from other datasets. During the feature mapping process, if these unique features are forced to be aligned like domain-invariant features to obtain domain independence, the generalization ability of the model will be reduced.

[0013] To address the above issues, the present invention takes into account the common features and unique features of ADHD data sets from multiple medical centers, extracts the common features of multiple data sets to align different data distributions, and extracts the unique features of each data set to enhance the generalization ability; in addition, an attention module is set to ensure the interpretability of the mapping features, and ADHD biomarkers are extracted based on this to enhance the medical interpretability of the model. Summary of the invention

[0014] The purpose of the present invention is to provide intelligent auxiliary diagnosis for attention deficit hyperactivity disorder (ADHD) and solve the problem that when the data to be diagnosed come from multiple different medical centers and are heterogeneous, the diagnostic accuracy is significantly reduced. To achieve the above purpose, the present invention provides an ADHD auxiliary diagnosis method based on domain generalization.

[0015] An embodiment of the present invention provides an attention deficit hyperactivity disorder auxiliary diagnosis method based on domain generalization, the method comprising:

[0016] Step 1: After preprocessing, K brain functional magnetic resonance imaging data sets with different probability distributions (each data set contains data of patients with attention deficit hyperactivity disorder and healthy people) are labeled (that is, the labeled data belongs to patients with attention deficit hyperactivity disorder or healthy people) to form labeled training data;

[0017] Step 2: Input the labeled training data into the shared attention network to extract the shared attention V;

[0018] Step 3: The extracted shared attention V is used as the weight of the training data, multiplied by the training data, and then input into the encoder network to extract the features of the training data;

[0019] Step 4: The features obtained in step 3 are input into the decoder network and the classification network respectively. The decoder network is used to convert the features back to the original data space, and the classification network is used for feature classification;

[0020] Step 5: Optimize the encoder and decoder networks through the reconstruction loss function, optimize the classification network through the classification cross entropy loss function, and align the data distribution of different data sets through the maximum mean difference loss function. When the changes in the above loss functions are all less than the predetermined threshold, the trained encoder, decoder, and classification network are obtained;

[0021] Step 6: Input the training data of the K data sets mentioned in step 1 into K unique attention networks respectively, and extract the unique attention Z of each data set i (i∈{1, 2, ..., K}) to constrain the shared attention V;

[0022] Step 7: Make orthogonal constraints between the K unique attentions Z extracted in step 6, and between the unique attention Z and the shared attention V extracted in step 1, to ensure that the extracted unique attention and shared attention features do not overlap;

[0023] Step 8: Apply the unique attention Z for each dataset extracted in step 6 i Add it to the shared attention V extracted in step 2 to get the weight W i (i.e. W i=Z i +V,i∈{1,2,...,K});

[0024] Step 9: Set the weight W i (i∈{1, 2, ..., K}), after multiplying with the data of the i-th training data set, it is input into the i-th classification network for classification;

[0025] Step 10: For each classification network described in step 9, optimize it through the classification cross entropy loss function to obtain K trained classification networks;

[0026] Step 11: After preprocessing, the K+1th brain functional magnetic resonance imaging data set (the data set contains data of patients with attention deficit hyperactivity disorder and healthy people) is used as a test data set, and is input into the final network structure model obtained from steps 2 to 5, and the classification result of the K+1th data set is output (that is, the test data is patients with attention deficit hyperactivity disorder or healthy people);

[0027] Step 12: Based on the output of the shared attention network in step 11, determine the significant brain regions of patients with attention deficit hyperactivity disorder in the test data.

[0028] In a preferred embodiment of the present invention, the data preprocessing of step 1 and step 11 includes removing various types of noise from functional magnetic resonance imaging (BOLD-fMRI) time series data using blood oxygen level dependence (BOLD) method through data processing means to obtain different biological signal data, including amplitude of low frequency fluctuations (ALFF), regional homogeneity (ReHo) and functional connectivity (FC).

[0029] In a preferred embodiment of the present invention, the shared attention network of step 2 includes a multi-layer fully connected structure, each fully connected structure includes a fully connected layer and an activation function layer in sequence, wherein the fully connected layer uses a 2-norm regularization constraint to control the model complexity and prevent overfitting; the final output layer of the shared attention network uses a Softmax function.

[0030] In a preferred embodiment of the present invention, the encoding network in step 3 includes a multi-layer fully connected structure, each of which includes a fully connected layer and an activation function layer in sequence, wherein the fully connected layer uses 2-norm regularization to constrain the weight of the layer, and the constraint strength is 0.01.

[0031] In a preferred embodiment of the present invention, the decoding network of step 4 includes multiple layers of fully connected layers, wherein the weights of the fully connected layers are initialized using the he_normal method; the classification network of step 4 includes multiple layers of fully connected layers, each fully connected structure includes a fully connected layer and an activation function layer in sequence, and the final output of the classification network uses a Softmax function.

[0032] In a preferred embodiment of the present invention, the unique attention network of step 6 has the same structure as the shared attention network of step 2; the classification network of step 9 has the same structure as the classification network of step 4.

[0033] In a preferred embodiment of the present invention, the method of obtaining the significant brain regions in step 12 is as follows:

[0034] According to the output of the common attention network in step 11, the attention value corresponding to each brain region is obtained, the attention values ​​are sorted from large to small, and the brain regions corresponding to the first 5 attention values ​​are selected as significant brain regions. The higher the attention value corresponding to the brain region, the more the model relies on the brain region when making decisions, that is, the more significant the role of the brain region in judging whether the disease is serious or not, and thus it is defined as a significant brain region.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] (1) The present invention aims to solve the problem that the diagnostic accuracy rate of attention deficit hyperactivity disorder is significantly reduced due to the heterogeneity of input data in intelligent auxiliary diagnosis of attention deficit hyperactivity disorder. The domain generalization method is used to extract the common features and unique features of data with different distributions, thereby improving the generalization ability of the model.

[0037] (2) In order to solve the problem of lack of medical interpretability of model-extracted features, the present invention uses the attention mechanism and designs an attention network to identify the features of brain regions that are key to the diagnosis of attention deficit hyperactivity disorder, thereby obtaining significant brain regions that play a key role in the diagnosis of attention deficit hyperactivity disorder, enhancing the interpretability of the features and enabling the output results of the model to provide a basis for doctors to make reliable diagnoses.

[0038] (3) The test results of the present invention on the ADHD-200 data set show that when there is more training data, a classification accuracy of 72.8% is obtained, and the significant brain regions obtained are consistent with the existing medical results of attention deficit hyperactivity disorder. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a schematic diagram of the steps of an attention deficit hyperactivity disorder auxiliary diagnosis method based on domain generalization provided by an embodiment of the present invention;

[0040] Figure 2It is a core network composition diagram in an attention deficit hyperactivity disorder auxiliary diagnosis method based on domain generalization provided by an embodiment of the present invention;

[0041] Figure 3 This is a salient brain region map of patients with attention deficit hyperactivity disorder obtained on a test data set according to an embodiment of the present invention. DETAILED DESCRIPTION

[0042] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0043] See also Figure 1 The present invention provides an auxiliary diagnosis method for attention deficit hyperactivity disorder based on domain generalization, comprising the following steps:

[0044] Step 1: After preprocessing, K brain functional magnetic resonance imaging data sets with different probability distributions (each data set contains data of patients with attention deficit hyperactivity disorder and healthy people) are labeled (that is, the labeled data belongs to patients with attention deficit hyperactivity disorder or healthy people) to form labeled training data, denoted as Where N is the total number of training data, x i represents the i-th training data, x i ∈R d (d represents the data dimension), y i Represents x i The label of y i ∈{1, 2, ..., C}, C represents the total number of categories of training data; at the same time, D tr ={D 1 , D 2 , …, D K} indicates that the training data includes K data sets. In this embodiment, C=2, K=3, x i The dimension is 90×1, where 90 represents the number of brain regions.

[0045] Specifically, during the collection of brain functional magnetic resonance imaging data, various factors such as the subject's head movement, external magnetic field, scanning equipment instability, and machine noise may cause different brain area activations to be recorded, thus affecting the accuracy of the research conclusions. Therefore, before analyzing brain functional magnetic resonance imaging data, the data must be preprocessed. The specific process mainly includes:

[0046] (1) Discarding initial data: Due to the inhomogeneity of the scanner magnetic field and the influence of the subjects’ initial adaptation stage, the signals at the first four time points are not stable enough and the data credibility is low, so they need to be removed.

[0047] (2) Head motion correction: First, the acquired fMRI image sequence is corrected for head motion to eliminate the effects caused by the subject's head movement during the scan. This is usually achieved by spatially registering the images at each time point in the sequence.

[0048] (3) Time correction: If there is a time offset or misalignment during the acquisition, the time series needs to be corrected to ensure that the data at each time point corresponds to the same time period.

[0049] (4) Spatial smoothing: To reduce the noise in the image and improve the signal-to-noise ratio, the data needs to be smoothed, which is achieved by weighted averaging the values ​​of each voxel in the brain image and its surrounding neighborhood. Commonly used smoothing methods include Gaussian filtering, which retains the spatial resolution of the image and reduces noise. The main purpose of spatial smoothing is to enhance the stability of the image, making subsequent statistical analysis more reliable.

[0050] (5) Spatial normalization: Align the scanned brain anatomy with the standard brain image space to compare data between different subjects or with previous research results. The registration is performed using the Montreal Neurological Institute (MNI) space at a 4×4×4 voxel resolution.

[0051] After removing various types of noise in the data through the above data processing methods, different biological signal data are obtained, including Amplitude of Low Frequency Fluctuations (ALFF), Regional Homogeneity (ReHo) and Functional Connectivity (FC), among which:

[0052] ALFF data were obtained by Fourier transforming the average time series of each extracted brain region to obtain a frequency domain representation:

[0053]

[0054] Where X i (f) is the Fourier transform result of the time series of the i-th brain region at frequency f. Then the power spectrum of each brain region in a specific low-frequency range (0.01-0.08Hz) is calculated, and the square root is taken to obtain the ALFF value of the brain region:

[0055]

[0056] ALFF i represents the ALFF value of the ith brain region.

[0057] ReHo data are obtained by calculating Kendall's coefficient of coordination for each pair of voxels in a certain brain region. This can be done using the following formula:

[0058]

[0059] Here, x i (t) and x j (t) is the signal of two different voxels at time t, and t' is another time point, sgn represents the sign function, N is the total number of voxels in a certain brain region, and the value of W is calculated over all voxel pairs. The ReHo value for the entire brain region can be obtained by averaging the above Kendall's coordination coefficient. This represents the average synchronization or consistency of the time series between all voxels in the brain region.

[0060] The construction method of FC data is as follows: first, the preprocessed data is divided into different brain regions according to the AAL-116 (Automated Anatomical Labeling-116, AAL-116) atlas. Each brain region is regarded as a node of the brain, and time series data is extracted from each brain region. Usually each brain region contains multiple voxels, and the average signal is generally used:

[0061]

[0062] in, is the average signal of the ith brain region at time t, X ik (t) is the signal of the kth voxel in the i-th brain region at time t, N i is the total number of voxels in the ith brain region. The time series correlation between brain regions is calculated to form a connection matrix C. The correlation is usually calculated using the Pearson correlation coefficient:

[0063]

[0064] The obtained correlation coefficient r ij Fill in the corresponding position of the connection matrix C, where C ij represents the connection strength between the i-th brain region and the j-th brain region.

[0065] In this embodiment, K brain functional magnetic resonance imaging data sets with different probability distributions (each data set contains data of patients with attention deficit hyperactivity disorder and healthy people) are preprocessed to obtain the ALFF biological signal of the brain of each sample (i.e., each subject) in the data set; then, the data is labeled (i.e., the labeled data belongs to patients with attention deficit hyperactivity disorder or healthy people) to form labeled training data, which is recorded as Where N is the total number of training data, xi represents the i-th training data, x i ∈R d (d represents the data dimension), y i Represents x i The label of y i ∈{1, 2, ..., C}, C represents the total number of categories of training data; at the same time, D tr ={D 1 , D 2 , …, D K} indicates that the training data includes K data sets. In this embodiment, C=2, K=3, x i The dimension is 90×1, where 90 represents the number of brain regions.

[0066] Step 2: Input the labeled training data into the shared attention network to extract the shared attention V.

[0067] Specifically, if Figure 2 As shown, the shared attention network adopts a network structure of 90×90Dense+ReLU activation function+90×90Dense+softmax activation function, where 90 represents the output attention dimension, which should be consistent with the number of brain regions in the input data. In this embodiment, each individual in the input data corresponds to 90 brain regions, and a 2-norm regularization constraint is used in each Dense layer to control the complexity of the model and prevent overfitting; the output dimension of V is 90×1.

[0068] Step 3: Use the extracted shared attention V as the weight of the training data, multiply it with the training data, and input it into the encoder network.

[0069] Specifically, the training data x i (dimension 90×1) is multiplied by the corresponding elements of the shared attention V (dimension 90×1) and then input into the encoder network.

[0070] The encoder network is Figure 2 As shown in the figure, a network structure of 90×50Dense+ReLU activation function+50×50Dense+ReLU activation function+50×50Dense is used, and a 2-norm regularization constraint is used in each Dense layer with a constraint strength of 0.01. The training data is mapped to the feature space through a multi-layer fully connected network to obtain high-level features.

[0071] Step 4: The high-level features obtained in step 3 are input into the decoder network and the classification network respectively.

[0072] Specifically, if Figure 2As shown in the figure, the decoder network adopts a 50×90Dense network structure, whose dimension is consistent with the encoder output dimension, which is used to convert the features back to the original data space. When the network starts training, the he_normal method is used to initialize the weights of the fully connected layer. The classification network adopts a 50×2Dense network structure. The Softmax activation function is used in this Dense layer to convert the output into a probability value for feature classification.

[0073] Step 5: Optimize the encoder and decoder networks through the reconstruction loss function, optimize the classification network through the classification cross entropy loss function, and align the data distribution of different data sets through the maximum mean difference loss function. When the changes in the above loss functions are all less than the predetermined threshold, the trained encoder, decoder, and classification network are obtained.

[0074] Specifically, by reconstructing the loss function Figure 1 The encoder and decoder networks in branch 1 are optimized, and the reconstruction loss function expression is:

[0075]

[0076] in Indicates the corresponding x i The encoder output of , ||·||2 represents the 2-norm operation.

[0077] Through the classification cross entropy loss function Figure 1 The classification network in branch 1 is optimized, and the cross entropy loss function expression is:

[0078]

[0079] in Represents the classification network for x i The predicted value of the classification.

[0080] Through the maximum mean difference loss function Figure 1 The data distribution of different data sets in branch 1 is aligned to shorten the distance between different data sets. First, calculate the data set D i and D j The expression of the maximum mean discrepancy (MMD) between is:

[0081]

[0082] Among them, i and j are the indices of the data set, and φ(·) is the mapping function from the original space to the feature space. Secondly, the maximum mean difference loss is calculated by taking the average of the MMD distances of the two domain pairs, that is, the maximum mean difference loss function is:

[0083]

[0084] When the changes in reconstruction loss, cross entropy loss, and optimal mean difference loss are all less than a predetermined threshold, the trained encoder, decoder, and classification network are obtained.

[0085] Step 6: Input the training data of the K data sets mentioned in step 1 into K unique attention networks respectively, and extract the unique attention Z of each data set i (i∈{1, 2, ..., K}), to constrain the shared attention V.

[0086] Specifically, if Figure 1 The second branch shown in Figure 2, where each unique attention network is composed of Figure 2 As shown, the network structure includes 90×90Dense+ReLU activation function+90×90Dense+softmax activation function, where 90 represents the output attention dimension, which should be consistent with the number of brain regions in the input data. In this embodiment, each individual in the input data corresponds to 90 brain regions, and a 2-norm regularization constraint is used in each Dense layer to control the model complexity and prevent overfitting; Z i The output dimension is 90×1.

[0087] Step 7: Make orthogonal constraints between the K unique attentions Z extracted in step 6, and between the unique attention Z and the shared attention V extracted in step 1, to ensure that the extracted unique attention and shared attention features do not overlap.

[0088] Specifically, in order to ensure the uniqueness of the unique attention extracted from each data set, it is required that there is no intersection between the unique attention and between the unique attention and the shared attention. Therefore, V and Z can be constrained to be orthogonal to each other, and R = {Z 1 , Z 2 , …, Z K , V}, then the orthogonal loss is:

[0089]

[0090] Where <·> represents the inner product of two vectors, and |·| represents taking the absolute value of each element. Through orthogonal constraints, it is ensured that the extracted unique attention and shared attention do not overlap.

[0091] Step 8: Apply the unique attention Z for each dataset extracted in step 6 i Add it to the shared attention V extracted in step 2 to get the weight W i (i.e. W i =Z i +V,i∈{1,2,…,K}).

[0092] Step 9: Set the weight W i (i∈{1, 2, ..., K}), multiplied by the data of the i-th training data set, and then input into the i-th classification network for classification.

[0093] Specifically, if Figure 2 As shown in the figure, each classification network adopts the network structure of 90×50Dense+ReLU activation function+50×50Dense+ReLU activation function+50×50Dense+50×2Dense+softmax activation function, in which the 2-norm regularization is used to constrain the weight value of the Dense layer with a constraint strength of 0.01. The last Dense layer uses the Softmax activation function to convert the output into a probability value.

[0094] Step 10: For each classification network described in step 9, optimize it through the classification cross entropy loss function to obtain K trained classification networks.

[0095] Specifically, the classification cross entropy loss function is:

[0096]

[0097] in represents the cross entropy loss of each classification network in branch 2, which is calculated in the same way as the classification cross entropy loss function described in step 5.

[0098] Step 11: After preprocessing, the K+1th brain functional magnetic resonance imaging data set (the data set contains data of patients with attention deficit hyperactivity disorder and healthy people) is used as a test data set, and input into the final network structure model obtained from steps 2 to 5, and output the classification result of the K+1th data set (that is, the test data is patients with attention deficit hyperactivity disorder or healthy people).

[0099] Specifically, the K+1th brain functional magnetic resonance imaging dataset is subjected to the same data preprocessing process as step 1 to obtain the ALFF biological signal data of each sample in the test set, and the test dataset is defined. Where M is the total number of test samples, represents the i-th test sample, (d represents the data dimension), express , and C represents the total number of categories of test samples, C = 2. In the embodiment of the present invention, K = 3, and the leave-one-out cross validation is adopted, that is, 3 data sets are selected from 4 data sets for training in each round, and one data set is used as test data, and a total of 4 rounds are performed until all 4 data sets are used as test data, and finally the average classification accuracy of the 4 rounds is calculated.

[0100] Step 12: Based on the output of the shared attention network in step 11, determine the significant brain regions of patients with attention deficit hyperactivity disorder in the test data.

[0101] Specifically, according to the output of the shared attention network in step 11, the attention value corresponding to each brain region is obtained, and the generative expression of shared attention is expressed as:

[0102] V = Softmax(w(Re LU(w(V init ))))

[0103] Where V init represents the initial value of V, w represents the linear transformation of the fully connected layer, ReLU and Softmax are activation functions, and it can be seen that V is between 0 and 1. Sort the attention values ​​V from large to small, and select the brain regions corresponding to the first 5 attention values ​​as significant brain regions. The higher the attention value corresponding to the brain region, the more the model relies on the brain region when making decisions, that is, the more significant the role of the brain region in judging whether the disease is serious or not, so it is defined as a significant brain region.

[0104] So far, according to the above-provided embodiment of an ADHD auxiliary diagnosis method based on domain generalization, the classification performance obtained includes accuracy, specificity, sensitivity, and AUC, as shown in Table 1. It can be seen that compared with other data sets, the embodiment of the present invention achieves the highest accuracy of 72.8% on the NI data set.

[0105] Table 1 Classification performance obtained by the embodiments of the present invention on different test data sets

[0106]

[0107] Table 2 Significant brain regions of patients with attention deficit hyperactivity disorder obtained on the test data set according to the embodiment of the present invention

[0108]

[0109] The reason why the highest accuracy is obtained on the NI dataset is that when the leave-one-out cross-validation method is adopted, when every three datasets are used as training sets, the number of the NYU+Peking+KKI datasets is the largest, and the model training is more sufficient, so that the generalization ability of the embodiment of the present invention is improved; on the contrary, the accuracy of the embodiment of the present invention on NYU is the worst. This is because when the NI+Peking+KKI datasets are used as training sets, the number is the smallest, which makes the model training insufficient and leads to poor accuracy.

[0110] The highest accuracy was obtained on the NI dataset. Therefore, step 12 was implemented on the NI dataset to extract the significant brain regions of patients with attention deficit hyperactivity disorder. The results are shown in Table 2. Figure 3 The specific locations of these brain regions are given, and the 5 significant brain regions selected are compared with the existing ADHD literature research results, which are consistent with the existing research results. First, these 5 brain regions are analyzed from the perspective of brain lobes. These brain regions are mainly located in the parietal lobe, frontal lobe and occipital lobe. These brain lobes are related to attention control and emotion regulation. Related studies have shown that the parietal (front) motor lobe is related to action intention and behavioral awareness. Damage to this lobe often leads to difficulty in autonomous control of behavior, and studies have shown that the white matter volume of the parietal lobe of ADHD patients is significantly reduced. The frontal lobe is related to functions such as attention, emotion regulation, and behavioral cognition. ADHD patients also show atypical frontal white matter microstructure. The occipital lobe is responsible for the perception and processing of visual information, and ADHD girls show delayed occipital lobe development.

[0111] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for assisting diagnosis of attention deficit hyperactivity disorder based on domain generalization, characterized in that The steps include: Step 1: After preprocessing, K brain functional magnetic resonance imaging data sets with different probability distributions (each data set contains data of patients with attention deficit hyperactivity disorder and healthy people) are labeled (that is, the labeled data belongs to patients with attention deficit hyperactivity disorder or healthy people) to form labeled training data; Step 2: Input the labeled training data into the shared attention network to extract the shared attention V; Step 3: The extracted shared attention V is used as the weight of the training data, multiplied by the training data, and then input into the encoder network to extract the features of the training data; Step 4: The features obtained in step 3 are input into the decoder network and the classification network respectively. The decoder network is used to convert the features back to the original data space, and the classification network is used for feature classification; Step 5: Optimize the encoder and decoder networks through the reconstruction loss function, optimize the classification network through the classification cross entropy loss function, and align the data distribution of different data sets through the maximum mean difference loss function. When the changes in the above loss functions are all less than the predetermined threshold, the trained encoder, decoder, and classification network are obtained; Step 6: Input the training data of the K data sets mentioned in step 1 into K unique attention networks respectively, and extract the unique attention Z of each data set i (i∈{1, 2, ..., K}) to constrain the shared attention V; Step 7: Make orthogonal constraints between the K unique attentions Z extracted in step 6, and between the unique attention Z and the shared attention V extracted in step 1, to ensure that the extracted unique attention and shared attention features do not overlap; Step 8: Apply the unique attention Z for each dataset extracted in step 6 i Add it to the shared attention V extracted in step 2 to get the weight W i (i.e. W i =Z i +V,i∈{1,2,...,K}); Step 9: Set the weight W i (i∈{1, 2, ..., K}), after multiplying with the data of the i-th training data set, it is input into the i-th classification network for classification; Step 10: For each classification network described in step 9, optimize it through the classification cross entropy loss function to obtain K trained classification networks; Step 11: After preprocessing, the K+1th brain functional magnetic resonance imaging data set (the data set contains data of patients with attention deficit hyperactivity disorder and healthy people) is used as a test data set, and is input into the final network structure model obtained from steps 2 to 5, and the classification result of the K+1th data set is output (that is, the test data is patients with attention deficit hyperactivity disorder or healthy people); Step 12: Based on the output of the shared attention network in step 11, determine the significant brain regions of patients with attention deficit hyperactivity disorder in the test data.

2. According to claim 1, a method for assisting diagnosis of attention deficit hyperactivity disorder based on domain generalization, characterized in that: The data preprocessing of step 1 and step 11 includes removing various types of noise from functional magnetic resonance imaging (BOLD-fMRI) time series data using blood oxygen level dependence (BOLD) by data processing methods to obtain different biological signal data, including low frequency fluctuation amplitude (Amplitude of Low Frequency Fluctuations, ALFF), local consistency (Regional Homogeneity, ReHo) and functional connectivity (Functional Connectivity, FC).

3. The method for assisting diagnosis of attention deficit hyperactivity disorder based on domain generalization according to claim 1, characterized in that: The shared attention network of step 2 includes a multi-layer fully connected structure, each of which includes a fully connected layer and an activation function layer in sequence, wherein the fully connected layer uses a 2-norm regularization constraint to control model complexity and prevent overfitting; the final output layer of the shared attention network uses a Softmax function.

4. The method for assisting diagnosis of attention deficit hyperactivity disorder based on domain generalization according to claim 1, characterized in that: The encoding network in step 3 includes a multi-layer fully connected structure, each of which includes a fully connected layer and an activation function layer in sequence, wherein the fully connected layer uses 2-norm regularization to constrain the weight of the layer, and the constraint strength is 0.

01.

5. The method for assisting diagnosis of attention deficit hyperactivity disorder based on domain generalization according to claim 1, characterized in that: The decoding network of step 4 includes multiple layers of fully connected layers, wherein the weights of the fully connected layers are initialized using the he_normal method; the classification network of step 4 includes multiple layers of fully connected layers, each fully connected structure includes a fully connected layer and an activation function layer in sequence, and the final output of the classification network uses a Softmax function.

6. The method for assisting diagnosis of attention deficit hyperactivity disorder based on domain generalization according to claim 1, characterized in that: The unique attention network of step 6 has the same structure as the shared attention network of step 2 of claim 3; the classification network of step 9 has the same structure as the classification network of step 4 of claim 5.

7. The method for assisting diagnosis of attention deficit hyperactivity disorder based on domain generalization according to claim 1, characterized in that: The method for obtaining the significant brain regions in step 12 is as follows: According to the output of the shared attention network in step 11 of claim 1, an attention value corresponding to each brain region is obtained, the attention values ​​are sorted from large to small, and the brain regions corresponding to the first 5 attention values ​​are selected as significant brain regions; the higher the attention value corresponding to a brain region, the more the model relies on the brain region when making decisions, that is, the more significant the role of the brain region in judging whether the disease is serious or not, and thus it is defined as a significant brain region.

Citation Information

Cited By

  • Method, device and system for assisting in recognizing attention deficit hyperactivity disorder based on electrocardiogram signals

    CN119314657A