Intelligent autistic child diagnosis system based on resting state near infrared spectrum

Abstract

The invention discloses an intelligent autistic child diagnosis system based on a resting state near infrared spectrum. The intelligent autistic child diagnosis system comprises a data import module, a matrix calculation module, an intelligent diagnosis training module and an intelligent diagnosis prediction module. Through loading 52 channel resting state near infrared spectrum data, correlation coefficients of relative concentration changes of hemoglobin between channels are calculated after preprocessing, and a 52 * 52 functional connection matrix is constructed. An improved CNN-Attention neural network is adopted for training and prediction, the CNN part captures local channel correlation, the self-attention mechanism layer highlights important channel pairs, and finally the diagnosis probability that whether autistic children suffer from hyperactivity or not is output. The method is low in children adaptability requirement, efficient in diagnosis process, high in operation speed and high in classification accuracy, supports new data addition and model retraining, and provides a scientific and visual diagnosis result.

Description

Technical Field

[0001] The present invention relates to the technical field of near-infrared brain functional imaging, and in particular to an intelligent diagnosis system for autistic children based on resting-state near-infrared spectroscopy. Background Art

[0002] Autism Spectrum Disorder (ASD) is a group of neurodevelopmental disorders mainly characterized by social communication disorders, restricted interests or activities, and repetitive stereotyped behaviors. This disorder appears before the age of 3, and autistic children have significant difficulties in social interaction, communication skills, and behavioral patterns.

[0003] In addition, the probability of autistic children co-occurring with Attention Deficit Hyperactivity Disorder (ADHD) is very high, and the symptoms of autism and hyperactivity disorder highly overlap, which brings additional complexity to diagnosis and treatment.

[0004] Currently, the clinical diagnosis of autistic children and autistic children co-occurring with ADHD mainly relies on medical history, neuropsychological assessment, doctors' behavioral observation, and parents' interviews, etc., and the diagnosis is highly subjective.

[0005] In order to make up for the problem of strong subjectivity in current clinical diagnosis, electroencephalogram (EEG) and functional magnetic resonance imaging (fMRI) can be used to diagnose autistic children, but these two methods require a high degree of cooperation from autistic children and it is difficult to obtain an accurate diagnosis result. Summary of the Invention

[0006] The purpose of the present invention is to overcome the shortcomings and deficiencies of existing clinical diagnosis methods, and propose an intelligent diagnosis system for autistic children based on resting-state near-infrared spectroscopy, which can provide a highly interpretable and highly reliable diagnosis result while having a low requirement for the cooperation of autistic children.

[0007] To achieve the above purpose, the technical solution provided by the present invention is: an intelligent diagnosis system for autistic children based on resting-state near-infrared spectroscopy, including:

[0008] A data import module for loading 52-channel resting-state near-infrared spectroscopy data of children, which has two disease labels, namely autism and autism co-occurring with ADHD, and preprocessing the data to obtain 52-channel relative hemoglobin concentration change data of children;

[0009] A matrix calculation module for calculating the correlation coefficient of the relative hemoglobin concentration change data between channels of children to form a 52×52 functional connectivity matrix;

[0010] An intelligent diagnosis training module is used to input the functional connection matrix and disease labels of children into an improved CNN-Attention neural network for training. The structure of this neural network is improved for the functional connection matrix of autistic children. The first half is a CNN network, which is used to capture the local channel correlation patterns of the functional connection matrix and reduce the number of channel pairs, focusing on stronger correlations. The second half is a self-attention mechanism layer, which is used to highlight the more important channel pairs in the classification task;

[0011] An intelligent diagnosis prediction module is used to input the 52-channel resting-state near-infrared spectroscopy data of autistic children with unknown disease labels into a data import module and a matrix calculation module to obtain a functional connection matrix, and then input this functional connection matrix into the trained improved CNN-Attention neural network to obtain the diagnostic probability result of the co-occurrence of ADHD in this autistic child.

[0012] Furthermore, the data import module includes a data loading module and a data preprocessing module, where:

[0013] The data loading module loads the 52-channel resting-state near-infrared spectroscopy data of children from local, and this data has two disease labels, namely autism and autism co-occurring with ADHD;

[0014] The data preprocessing module calculates the original light intensity signal of the resting-state near-infrared spectroscopy data of each channel of children into optical density, and then converts it into hemoglobin relative concentration change data through the modified Beer-Lambert law, including the following steps:

[0015] 1) Calculate the optical density:

[0016]

[0017] In the formula, OD is the optical density, indicating the degree of light absorption; I 入 is the incident light intensity; I 透 is the transmitted light intensity;

[0018] 2) Calculate the relative concentration change of hemoglobin:

[0019] The calculation formula of the modified Beer-Lambert law is as follows:

[0020] OD = ∈·C·DPF·L + G

[0021] Where ∈ is the extinction coefficient, representing the ability of a substance to absorb light; DPF is the differential path length correction factor, representing how many times the actual path length of light in tissue is the straight-line length due to the scattering effect; C is the concentration of the light-absorbing substance; L is the path length through which the light passes; G is the total attenuation of the light intensity caused by other factors, representing the attenuation of the light intensity caused by other factors except the light-absorbing substance.

[0022] Therefore, the calculation formula for the relative change in hemoglobin concentration is as follows:

[0023]

[0024] Where ΔC is the data of the relative change in hemoglobin concentration; ΔOD is the difference in the change of optical density.

[0025] Then, perform motion artifact correction, filtering, baseline correction, and normalization on the data of the relative change in hemoglobin concentration.

[0026] Furthermore, the matrix calculation module is used to calculate the correlation coefficients of the relative change in hemoglobin concentration data between channels of children, form a 52×52 functional connectivity matrix, and perform Fisher Z transformation, including the following steps:

[0027] 1) Calculate the correlation coefficients of the relative change in hemoglobin concentration data between channels of children, form a 52×52 functional connectivity matrix, where the correlation coefficient uses the Pearson correlation coefficient:

[0028]

[0029] Where r ij is the Pearson correlation coefficient between the i-th channel and the j-th channel; x ik is the value of the i-th channel at the k-th observation; x jk is the value of the j-th channel at the k-th observation; is the average value of all observations of the i-th channel, that is is the average value of all observations of the j-th channel, that is n is the number of observations for each channel;

[0030] 2) Perform Fisher Z transformation on each correlation coefficient of the functional connectivity matrix to convert it into a matrix with an approximately normal distribution for statistical analysis:

[0031]

[0032] Where r is the correlation coefficient, and z r is the correlation coefficient after Fisher Z transformation.

[0033] Furthermore, the structure and function of each layer of the improved CNN-Attention neural network are as follows:

[0034] The first layer is the first convolutional layer, which contains 32 convolutional kernels of size 3×3, and the activation function is ReLU. It is used to extract local features of the input functional connectivity matrix. The formula is as follows:

[0035] X conv1 = ReLU(W conv1 * X + b conv1 )

[0036] In the formula, X conv1 is the output of the first convolutional layer, W conv1 is the weight matrix of the convolutional kernels of the first convolutional layer, * is the convolution operation, X is the input functional connectivity matrix, and b conv1 is the bias term of the first convolutional layer;

[0037] The second layer is the first max-pooling layer, with a pooling window size of 2×2. It is used to reduce the spatial dimension of the features extracted by the previous layer, reduce the computational amount, and retain important features at the same time. The formula is as follows:

[0038] X pool1 = MaxPool 2×2 (X conv1 )

[0039] In the formula, X pool1 is the output of the first max-pooling layer;

[0040] The third layer is the second convolutional layer, which contains 64 convolutional kernels of size 3×3, and the activation function is ReLU. It is used to further extract higher-level local features of the input functional connectivity matrix. The formula is as follows:

[0041] X conv2 = ReLU(W conv2 * X pool1 + b conv2 )

[0042] In the formula, X conv2 is the output of the second convolutional layer, W conv2 is the weight matrix of the convolutional kernels of the second convolutional layer, and b conv2 is the bias term of the second convolutional layer;

[0043] The fourth layer is the second max-pooling layer, with a pooling window size of 2×2. It is used to reduce the spatial dimension of the features extracted by the previous layer again. The formula is as follows:

[0044] X pool2 = MaxPool 2×2 (X conv2)

[0045] Wherein, X pool2 is the output of the second maximum pooling layer;

[0046] The fifth layer is the Flatten layer, which is used to flatten the multi-dimensional features into a one-dimensional feature vector for input into the fully connected layer. The formula is as follows:

[0047] X flat = Flatten(X pool2 )

[0048] Wherein, X flat is the output of the Flatten layer;

[0049] The sixth layer is the fully connected layer, which contains 128 neurons and the activation function is ReLU. It is used to further extract features and classify the one-dimensional feature vector. The formula is as follows:

[0050] X dense = ReLU(W dense X flat + b dense )

[0051] Wherein, X dense is the output of the fully connected layer, W dense is the weight matrix of the fully connected layer, and b dense is the bias term of the fully connected layer;

[0052] The seventh layer is the Dropout layer, and the proportion of Dropout is 0.5. It is used to randomly discard the outputs of some neurons to prevent overfitting. The formula is as follows:

[0053] X dropout = Dropout(X dense )

[0054] Wherein, X dropout is the output of the Dropout layer;

[0055] The eighth layer is the self-attention mechanism layer with a dimension of 128. It is used to weight the output of the Dropout layer to highlight important features. In the self-attention mechanism, Query represents the query intention for the input features, that is, the part to be concerned about, Key represents the key part of the input features, that is, which features are worthy of attention, and Value represents the actual content of the input features, that is, the information to be extracted. The calculation formulas for Query, Key, and Value are as follows:

[0056] Q = W Q X dropout

[0057] K = WK X dropout

[0058] V = W V X dropout

[0059] Wherein, Q, K, and V are matrices corresponding to Query, Key, and Value respectively, and W Q , W K , W V are learnable weight matrices corresponding to Q, K, and V respectively; The attention scores are calculated by the dot product of Q and K, and the formula is as follows:

[0060]

[0061] Wherein, A is the attention score matrix, Q and K are matrices corresponding to Query and Key respectively, and d k is the dimension of Key, which is used to scale the dot product result to avoid the problem of vanishing gradients caused by excessive numerical values. The Softmax function is used for normalization; According to the attention scores, V is weighted and aggregated to generate the final attention output, and the formula is as follows:

[0062] X attention = AV

[0063] Wherein, X attention is the attention output matrix, and V is the matrix corresponding to Value;

[0064] Finally, by directly adding the input, that is, the output of the Dropout layer, to the output, the network can learn the difference part between the input and the output, rather than directly learning the output itself, and thus calculate the residual connection, and the formula is as follows

[0065] X final = X dropout + γX attention

[0066] Wherein, X final is the final output feature, and γ is a learnable weight coefficient, which is used to control the influence degree of the attention output on the final result;

[0067] The 9th layer is the output layer, which is a fully connected layer with an output dimension of 2 and an activation function of Softmax, and is used to map the features to the class probability distribution, and the formula is as follows:

[0068] Y = Softmax(W out X final + b out )

[0069] Wherein, Y is the output of the output layer, and W outis the weight matrix of the output layer, X final is the final output feature, b out is the bias term of the output layer;

[0070] The loss function of the improved CNN - Attention neural network is the cross - entropy loss function, which calculates the cross - entropy loss between the probability distribution of the network output and the target probability distribution, and is applicable to multi - classification problems; the optimizer used is the Adam optimizer, which is an adaptive learning rate optimization algorithm; the evaluation metric is accuracy, that is, the proportion of the number of correctly predicted samples to the total number of samples.

[0071] Furthermore, to train the improved CNN - Attention neural network, the specific process is as follows: the input data passes through each layer of the neural network in turn, that is, after forward propagation, backward propagation is carried out. The process of backward propagation includes calculating the loss, that is, using the cross - entropy loss function to calculate the cross - entropy loss between the probability distribution of the network output and the target probability distribution, calculating the gradients, that is, calculating the gradients of the weight matrix and weight coefficients of the network according to the cross - entropy loss function, and updating the parameters, that is, using the Adam optimizer to update the weight matrix and weight coefficients of the network according to the calculated gradients; in each epoch, the network performs multiple forward propagations and backward propagations on the training data, continuously updating the weight matrix and weight coefficients to minimize the loss; at the end of each epoch, the accuracy is used as the evaluation metric to evaluate the performance of the network on the validation set.

[0072] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0073] 1. It has low requirements for the cooperation of autistic children, which is conducive to the smooth completion of diagnosis.

[0074] 2. It provides the diagnostic probability of whether autistic children co - suffer from ADHD, which is more scientific and intuitive.

[0075] 3. By using the improved CNN - Attention neural network, it has a fast running speed, high classification accuracy, and is convenient to add new data and retrain the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 is a schematic diagram of the relationship between the various modules of the system of the present invention.

[0077] Figure 2 is a flowchart of the training and prediction of the system of the present invention.

[0078] Figure 3 is a schematic diagram of the functional connection matrix.

[0079] Figure 4It is the structural diagram of the improved CNN-Attention neural network; in the figure, ASD represents children diagnosed with autism, and ASD&ADHD represents children diagnosed with autism co-occurring with ADHD. Specific implementation manners

[0080] The present invention will be further described in detail below in conjunction with embodiments and the accompanying drawings, but the implementation manners of the present invention are not limited thereto.

[0081] This embodiment discloses an intelligent diagnosis system for autistic children based on resting-state near-infrared spectroscopy, which is an intelligent diagnosis system for autistic children developed using the Python language and can run on Windows devices. The relationships between the various modules of the system are as Figure 1 shown, and the flowcharts of system training and prediction are as Figure 2 shown. It includes:

[0082] A data import module, which is used to load the 52-channel resting-state near-infrared spectroscopy data of children. This data has two disease labels, namely autism and autism co-occurring with ADHD, and preprocesses the data to obtain the 52-channel relative hemoglobin concentration change data of children;

[0083] A matrix calculation module, which is used to calculate the correlation coefficients of the relative hemoglobin concentration change data between channels of children to form a 52×52 functional connectivity matrix, as Figure 3 shown;

[0084] An intelligent diagnosis training module, which is used to input the functional connectivity matrix and disease labels of children into the improved CNN-Attention neural network for training. The structure of this neural network is improved for the functional connectivity matrix of autistic children. The first half is a CNN network, which is used to capture the local channel correlation patterns of the functional connectivity matrix and reduce the number of channel pairs, focusing on stronger correlations. The second half is a self-attention mechanism layer, which is used to highlight the more important channel pairs in the classification task;

[0085] An intelligent diagnosis prediction module, which is used to input the 52-channel resting-state near-infrared spectroscopy data of autistic children with unknown disease labels into the data import module and the matrix calculation module to obtain a functional connectivity matrix, and then input this functional connectivity matrix into the trained improved CNN-Attention neural network to obtain the diagnosis probability result of this autistic child co-occurring with ADHD.

[0086] Specifically, the data import module includes a data loading module and a data preprocessing module, where:

[0087] The data loading module loads the 52-channel resting-state near-infrared spectroscopy data of children from local, and this data has two disease labels, namely autism and autism co-occurring with attention deficit hyperactivity disorder;

[0088] The data preprocessing module calculates the original light intensity signal of the resting-state near-infrared spectroscopy data of each channel of children into optical density, and then converts it into hemoglobin relative concentration change data through the modified Beer-Lambert law, including the following steps:

[0089] 1) Calculate the optical density:

[0090]

[0091] In the formula, OD is the optical density, indicating the degree of light absorption; I 入 is the incident light intensity; I 透 is the transmitted light intensity;

[0092] 2) Calculate the relative concentration change of hemoglobin:

[0093] The calculation formula of the modified Beer-Lambert law is as follows:

[0094] OD = ∈·C·DPF·L + G

[0095] In the formula, ∈ is the extinction coefficient, indicating the ability of the substance to absorb light; DPF is the differential path length correction factor, indicating how many times the actual path length of light in the tissue is the straight-line length due to the scattering effect; C is the concentration of the light-absorbing substance; L is the path length that the light passes through; G is the total light intensity attenuation caused by other factors, indicating the light intensity attenuation caused by other factors except the light-absorbing substance;

[0096] Therefore, the calculation formula of the hemoglobin relative concentration change is as follows:

[0097]

[0098] In the formula, ΔC is the hemoglobin relative concentration change data; ΔOD is the difference in optical density change;

[0099] Then, motion artifact correction, filtering, baseline correction, and normalization are performed on the hemoglobin relative concentration change data.

[0100] Specifically, the matrix calculation module is used to calculate the correlation coefficient of the hemoglobin relative concentration change data between channels of children, form a 52×52 functional connectivity matrix, and perform Fisher Z transformation, including the following steps:

[0101] 1) Calculate the correlation coefficient of the hemoglobin relative concentration change data between channels of children, form a 52×52 functional connectivity matrix, where the correlation coefficient uses the Pearson correlation coefficient:

[0102]

[0103] Wherein, r ij is the Pearson correlation coefficient between the i-th channel and the j-th channel; x ik is the value of the i-th channel at the k-th observation; x jk is the value of the j-th channel at the k-th observation; is the average value of all observations of the i-th channel, that is is the average value of all observations of the j-th channel, that is n is the number of observations for each channel;

[0104] 2) Perform Fisher Z-transform on each correlation coefficient of the functional connectivity matrix to convert it into a matrix with an approximate normal distribution for statistical analysis:

[0105]

[0106] Wherein, r is the correlation coefficient, and z r is the correlation coefficient after Fisher Z-transform.

[0107] Specifically, as Figure 4 shown, the structure and function of each layer of the improved CNN-Attention neural network are as follows:

[0108] The first layer is the first convolutional layer, which contains 32 convolutional kernels of size 3×3, and the activation function is ReLU, which is used to extract the local features of the input functional connectivity matrix. The formula is as follows:

[0109] X conv1 = ReLU(W conv1 * X + b conv1 )

[0110] Wherein, X conv1 is the output of the first convolutional layer, W conv1 is the weight matrix of the convolutional kernels of the first convolutional layer, * is the convolution operation, X is the input functional connectivity matrix, and b conv1 is the bias term of the first convolutional layer;

[0111] The second layer is the first max-pooling layer, and the pooling window size is 2×2, which is used to reduce the spatial dimension of the features extracted by the previous layer, reduce the amount of calculation, and retain important features at the same time. The formula is as follows:

[0112] X pool1 = MaxPool 2×2 (X conv1 )

[0113] Wherein, X pool1 is the output of the first max pooling layer;

[0114] The third layer is the second convolutional layer, which contains 64 convolutional kernels of size 3×3, and the activation function is ReLU, which is used to further extract higher-level local features of the input functional connection matrix. The formula is as follows:

[0115] X conv2 = ReLU(W conv2 * X pool1 + b conv2 )

[0116] Wherein, X conv2 is the output of the second convolutional layer, W conv2 is the weight matrix of the convolutional kernels of the second convolutional layer, and b conv2 is the bias term of the second convolutional layer;

[0117] The fourth layer is the second max pooling layer, and the pooling window size is 2×2, which is used to reduce the spatial dimension of the features extracted by the previous layer again. The formula is as follows:

[0118] X pool2 = MaxPool 2×2 (X conv2 )

[0119] Wherein, X pool2 is the output of the second max pooling layer;

[0120] The fifth layer is the Flatten layer, which is used to flatten the multi-dimensional features into a one-dimensional feature vector for input into the fully connected layer. The formula is as follows:

[0121] X flat = Flatten(X pool2 )

[0122] Wherein, X flat is the output of the Flatten layer;

[0123] The sixth layer is the fully connected layer, which contains 128 neurons, and the activation function is ReLU, which is used to further extract features and classify the one-dimensional feature vector. The formula is as follows:

[0124] X dense = ReLU(W dense X flat + b dense )

[0125] Wherein, X dense is the output of the fully connected layer, W dense is the weight matrix of the fully connected layer, and b dense is the bias term of the fully connected layer;

[0126] The 7th layer is the Dropout layer with a Dropout ratio of 0.5, which is used to randomly discard the outputs of some neurons to prevent overfitting. The formula is as follows:

[0127] X dropout = Dropout(X dense )

[0128] In the formula, X dropout is the output of the Dropout layer;

[0129] The 8th layer is the self-attention mechanism layer with a dimension of 128, which is used to weight the output of the Dropout layer to highlight important features. In the self-attention mechanism, Query represents the query intention for the input features, that is, the part to be focused on, Key represents the key part of the input features, that is, which features are worthy of attention, and Value represents the actual content of the input features, that is, the information to be extracted. The calculation formulas for Query, Key, and Value are as follows:

[0130] Q = W Q X dropout

[0131] K = W K X dropout

[0132] V = W V X dropout

[0133] In the formula, Q, K, and V are the matrices corresponding to Query, Key, and Value respectively, and W Q , W K , W V are the learnable weight matrices corresponding to Q, K, and V respectively. The attention scores are calculated by the dot product of Q and K. The formula is as follows:

[0134]

[0135] In the formula, A is the attention score matrix, Q and K are the matrices corresponding to Query and Key respectively, d k is the dimension of Key, which is used to scale the dot product result to avoid the problem of gradient disappearance caused by too large numerical values. The Softmax function is used for normalization. According to the attention scores, V is weighted and aggregated to generate the final attention output. The formula is as follows:

[0136] X attention = AV

[0137] In the formula, X attention is the attention output matrix, and V is the matrix corresponding to Value;

[0138] Finally, by directly adding the input, which is the output of the Dropout layer, to the output, the network can learn the differential part between the input and the output instead of directly learning the output itself, thereby calculating the residual connection. The formula is as follows

[0139] X final = X dropout + γX attention

[0140] In the formula, X final is the final output feature, and γ is a learnable weight coefficient used to control the influence degree of the attention output on the final result;

[0141] The 9th layer is the output layer, which is a fully connected layer with an output dimension of 2 and an activation function of Softmax, and is used to map the features to the class probability distribution. The formula is as follows:

[0142] Y = Softmax(W out X final + b out )

[0143] In the formula, Y is the output of the output layer, W out is the weight matrix of the output layer, X final is the final output feature, and b out is the bias term of the output layer;

[0144] The loss function of the improved CNN - Attention neural network is the cross - entropy loss function, which calculates the cross - entropy loss between the probability distribution of the network output and the target probability distribution, and is applicable to multi - classification problems; the optimizer used is the Adam optimizer, which is an adaptive learning rate optimization algorithm; the evaluation metric is accuracy, that is, the proportion of the number of correctly predicted samples to the total number of samples.

[0145] To train the improved CNN - Attention neural network, the specific process is as follows: The input data passes through each layer of the neural network in turn, that is, after forward propagation, backward propagation is performed. The process of backward propagation includes calculating the loss, that is, using the cross - entropy loss function to calculate the cross - entropy loss between the probability distribution of the network output and the target probability distribution, calculating the gradients, that is, calculating the gradients of the weight matrix and weight coefficient of the network according to the cross - entropy loss function, and updating the parameters, that is, using the Adam optimizer to update the weight matrix and weight coefficient of the network according to the calculated gradients; in each epoch, the network performs multiple forward propagations and backward propagations on the training data, continuously updating the weight matrix and weight coefficient to minimize the loss; at the end of each epoch, the accuracy is used as the evaluation metric to evaluate the performance of the network on the validation set.

[0146] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention shall be equivalent substitution methods and are all included in the protection scope of the present invention.

Claims

1. An intelligent diagnosis system for autistic children based on resting-state near-infrared spectroscopy, characterized in that, Including: A data import module for loading 52-channel resting-state near-infrared spectroscopy data of children, which has two disease labels, namely autism and autism co-occurring with attention deficit hyperactivity disorder (ADHD), and preprocessing the data to obtain the relative concentration change data of hemoglobin in 52 channels of children; A matrix calculation module for calculating the correlation coefficients of the relative concentration change data of hemoglobin between channels of children to form a 52×52 functional connectivity matrix; An intelligent diagnosis training module for training by inputting the functional connectivity matrix and disease labels of children into an improved CNN-Attention neural network. The structure of this neural network is improved for the functional connectivity matrix of autistic children. The first half is a CNN network for capturing the local channel correlation patterns of the functional connectivity matrix and reducing the number of channel pairs, focusing on stronger correlations. The second half is a self-attention mechanism layer for highlighting the more important channel pairs in the classification task; An intelligent diagnosis prediction module for inputting the 52-channel resting-state near-infrared spectroscopy data of autistic children with unknown disease labels into the data import module and the matrix calculation module to obtain a functional connectivity matrix, and then inputting this functional connectivity matrix into the trained improved CNN-Attention neural network to obtain the diagnostic probability result of the autistic child co-occurring with ADHD.

2. The intelligent diagnosis system for autistic children based on resting-state near-infrared spectroscopy according to claim 1, wherein: The data import module includes a data loading module and a data preprocessing module, where: The data loading module loads the 52-channel resting-state near-infrared spectroscopy data of children from local, which has two disease labels, namely autism and autism co-occurring with ADHD; The data preprocessing module calculates the original light intensity signal of the resting-state near-infrared spectroscopy data of each channel of children into optical density, and then converts it into the relative concentration change data of hemoglobin through the modified Beer-Lambert law, including the following steps: 1) Calculate the optical density: Where OD is the optical density, representing the degree of light absorption; I 入 is the incident light intensity; I 透 is the transmitted light intensity; 2) Calculate the relative concentration change of hemoglobin: The calculation formula of the modified Beer-Lambert law is as follows: OD = ∈·C·DPF·L + G In the formula, ∈ is the extinction coefficient, representing the ability of the substance to absorb light; DPF is the differential path length correction factor, representing how many times the actual path length of light in the tissue is the straight-line length due to the scattering effect; C is the concentration of the light-absorbing substance; L is the path length that the light passes through; G is the total light intensity attenuation caused by other factors, representing the light intensity attenuation caused by other factors except the light-absorbing substance; Therefore, the calculation formula of the relative concentration change of hemoglobin is as follows: In the formula, ΔC is the relative concentration change data of hemoglobin; ΔOD is the difference in optical density change; Then, perform motion artifact correction, filtering, baseline correction, and normalization on the relative concentration change data of hemoglobin.

3. The intelligent diagnosis system for autistic children based on resting-state near-infrared spectroscopy according to claim 1, wherein: The matrix calculation module is used to calculate the correlation coefficients of the relative concentration change data of hemoglobin between channels of children to form a 52×52 functional connectivity matrix, and perform Fisher Z transformation, including the following steps: 1) Calculate the correlation coefficients of the relative hemoglobin concentration change data between channels of children to form a 52×52 functional connectivity matrix, where the correlation coefficient uses the Pearson correlation coefficient: where r ij is the Pearson correlation coefficient between the i-th channel and the j-th channel; x ik is the value of the i-th channel at the k-th observation; x jk is the value of the j-th channel at the k-th observation; is the average value of all observations of the i-th channel, that is is the average value of all observations of the j-th channel, that is n is the number of observations for each channel; 2) Perform Fisher Z-transform on each correlation coefficient of the functional connectivity matrix to convert it into a matrix with an approximate normal distribution for statistical analysis: where r is the correlation coefficient and z r is the correlation coefficient after Fisher Z transformation.

4. An intelligent diagnosis system for autistic children based on resting-state near-infrared spectroscopy according to claim 1, characterized in that: The structure and function of each layer of the improved CNN-Attention neural network are as follows: The first layer is the first convolutional layer, which contains 32 convolutional kernels of size 3×3, and the activation function is ReLU, which is used to extract local features of the input functional connectivity matrix. The formula is as follows: X conv1 = ReLU(W conv1 * X + b conv1 ) where X conv1 is the output of the first convolutional layer, W conv1 is the weight matrix of the convolutional kernel of the first convolutional layer, * represents the convolution operation, X is the input functional connection matrix, and b conv1 is the bias term of the first convolutional layer; The second layer is the first max-pooling layer, and the pooling window size is 2×2, which is used to reduce the spatial dimension of the features extracted by the previous layer, reduce the amount of calculation, and retain important features at the same time. The formula is as follows: X pool1 = MaxPool 2×2 (X conv1 ) where X pool1 is the output of the first max pooling layer; The third layer is the second convolutional layer, which contains 64 convolutional kernels of size 3×3, and the activation function is ReLU, which is used to further extract higher-level local features of the input functional connectivity matrix. The formula is as follows: X conv2 = ReLU(W conv2 * X pool1 + b conv2 ) where X conv2 is the output of the second convolutional layer, W conv2 is the weight matrix of the convolutional kernel of the second convolutional layer, b conv2 is the bias term of the second convolutional layer; The fourth layer is the second max-pooling layer, and the pooling window size is 2×2, which is used to reduce the spatial dimension of the features extracted by the previous layer again. The formula is as follows: X pool2 = MaxPool 2×2 (X conv2 ) where X pool2 is the output of the second maximum pooling layer; The fifth layer is the Flatten layer, which is used to flatten multi-dimensional features into one-dimensional feature vectors for input into the fully connected layer. The formula is as follows: X flat = Flatten(X pool2 ) where X flat is the output of the Flatten layer; The sixth layer is the fully connected layer, which contains 128 neurons, and the activation function is ReLU, which is used to further extract features and classify the one-dimensional feature vectors. The formula is as follows: X dense = ReLU(W dense X flat + b dense ) where X dense is the output of the fully connected layer, W dense is the weight matrix of the fully connected layer, and b dense is the bias term of the fully connected layer; The seventh layer is the Dropout layer, and the proportion of Dropout is 0.5, which is used to randomly discard the outputs of some neurons to prevent overfitting. The formula is as follows: X dropout = Dropout(X dense ) where X dropout is the output of the Dropout layer; The eighth layer is the self-attention mechanism layer with a dimension of 128, which is used to weight the output of the Dropout layer to highlight important features; in the self-attention mechanism, Query represents the query intention for the input features, that is, the part to be concerned about, Key represents the key part of the input features, that is, which features are worthy of attention, Value represents the actual content of the input features, that is, the information to be extracted. The calculation formulas for Query, Key, and Value are as follows: Q = W Q X dropout K = W K X dropout V = W V X dropout Wherein, Q, K, and V are matrices corresponding to Query, Key, and Value respectively, and W Q , W K , W V are learnable weight matrices corresponding to Q, K, and V respectively; The attention scores are calculated by the dot product of Q and K, and the formula is as follows: Where A is the attention score matrix, Q and K are the matrices corresponding to Query and Key respectively, and d k is the dimension of Key, which is used to scale the dot product result to avoid the problem of gradient disappearance caused by excessive numerical values. The Softmax function is used for normalization; According to the attention scores, V is weighted and aggregated to generate the final attention output. The formula is as follows: X attention = AV where X attention is the attention output matrix, and V is the matrix corresponding to Value; Finally, by directly adding the input, that is, the output of the Dropout layer, to the output, the network can learn the difference part between the input and the output instead of directly learning the output itself, thereby calculating the residual connection. The formula is as follows X final = X dropout + γX attention Where X final is the final output feature, and γ is a learnable weight coefficient used to control the influence degree of the attention output on the final result; The ninth layer is the output layer, which is a fully connected layer with an output dimension of 2 and an activation function of Softmax, which is used to map features to the category probability distribution. The formula is as follows: Y = Softmax(W out X final + b out ) where Y is the output of the output layer, W out is the weight matrix of the output layer, X final is the final output feature, b out is the bias term of the output layer; The loss function of the improved CNN-Attention neural network is the cross-entropy loss function, which calculates the cross-entropy loss between the probability distribution of the network output and the target probability distribution, and is applicable to multi-classification problems; the optimizer adopted is the Adam optimizer, which is an adaptive learning rate optimization algorithm; the evaluation metric is accuracy, that is, the proportion of the number of correctly predicted samples to the total number of samples.

5. The intelligent diagnosis system for autistic children based on resting-state near-infrared spectroscopy according to claim 4, wherein: Training the improved CNN-Attention neural network, the specific process is as follows: The input data passes through each layer of the neural network in turn, that is, after forward propagation, backpropagation is carried out. The process of backpropagation includes calculating the loss, that is, using the cross-entropy loss function to calculate the cross-entropy loss between the probability distribution of the network output and the target probability distribution, calculating the gradient, that is, calculating the gradients of the weight matrix and weight coefficients of the network according to the cross-entropy loss function, and updating the parameters, that is, using the Adam optimizer to update the weight matrix and weight coefficients of the network according to the calculated gradients; in each epoch, the network performs multiple forward propagations and backpropagations on the training data, continuously updating the weight matrix and weight coefficients to minimize the loss. At the end of each epoch, the accuracy is used as an evaluation metric to evaluate the performance of the network on the validation set.

Images