An intelligent diagnosis system for carpal tunnel syndrome based on DPCNN

Abstract

The carpal tunnel syndrome intelligent diagnosis system based on DPCNN of the present invention includes: collecting electromyographic data of relevant test items of multiple subjects; preprocessing each electromyographic data; splicing each electromyographic data after preprocessing of the subject into an electromyographic sample data, and diagnosing each electromyographic sample data according to the diagnosis report to construct a model data set; selecting the DPCNN structure to build a deep learning model; inputting each electromyographic sample data in the training set into the model, extracting features from the model for iterative training, and continuously adjusting hyperparameters in combination with the model training results until an optimal solution is obtained; performing multi-fold cross-validation training on the model, and constructing a confusion matrix according to the cross-validation results to comprehensively evaluate the model classification performance; using the model with the best performance of various indicators in the cross-validation as the optimal deep learning model; and inputting the electromyographic sample data to be predicted of a new subject into the optimal deep learning model to obtain the corresponding output intelligent diagnosis result.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an intelligent diagnosis system for carpal tunnel syndrome based on DPCNN (a deep pyramid convolutional neural network model, which is the first word-level widely effective deep text classification convolutional neural network in a strict sense). Background Art

[0002] Carpal tunnel syndrome (CTS) is caused by compression of the median nerve as it passes through the carpal tunnel formed by the flexor retinaculum. It is the most common peripheral compressive neuropathy and an important cause of work disability, with almost 3% of the general population affected by CTS. This carpal tunnel syndrome has various clinical manifestations ranging from mild pain to weakness or atrophy of thenar muscles, depending on the degree and duration of nerve compression. Currently, CTS is clinically subdivided into three subtypes: mild, moderate, and severe. Mild and moderate CTS usually use drug intervention, while severe CTS requires surgical treatment. Therefore, it is very important to correctly diagnose the severity of CTS and determine the appropriate treatment plan based on the severity.

[0003] Electromyography (EMG) is one of the standards for diagnosing peripheral nerve diseases. It plays a vital role in diagnosing CTS, and this technology is conducive to the classification of the severity of CTS. Traditional clinical CTS diagnosis and grading requires expert doctors to manually analyze multiple measurement values ​​in EMG detection of motor conduction and sensory conduction combined with clinical experience. This method of CTS diagnosis and grading is labor-intensive and has low efficiency in diagnosis and grading. In addition, the accuracy of manual classification results is positively correlated with the work experience of doctor experts, and accurate diagnosis and grading of CTS is difficult.

[0004] The diagnosis and grading of CTS is a multi-classification problem. With the advancement of computer technology and artificial intelligence technology in the 21st century, more and more studies have used this technology to achieve medical data classification. Machine learning is a subset of artificial intelligence technology, and many classic algorithms are of reference value for the diagnosis and grading of CTS. Professor Tsamis' team in Greece used a variety of machine learning algorithms such as support vector machine (SVM) and linear regression (LR) to analyze the EMG data of CTS patients and control volunteers. Among them, SVM performed well in the three-classification task of intelligently identifying healthy controls, mild / moderate CTS patients, and severe CTS patients. However, machine learning algorithms usually require complex feature engineering to achieve good classification performance. Researchers need to spend a lot of time to explore and analyze the original electromyographic signals and select appropriate features. In addition, machine learning models require a large amount of data for training, and have high requirements on the size of the data set. Therefore, the existing methods for diagnosing and grading CTS based on machine learning combined with EMG have complex data preprocessing and feature extraction processes. In addition, there is a risk of losing key information of EMG signals in the artificially selected features.

[0005] As a subset of machine learning, deep learning uses neural networks and hidden layers to automate feature engineering, greatly saving people's time and energy. Using raw data to train deep learning models to achieve tasks such as classification, regression, and clustering has made great progress in the fields of text, images, and audio. As the most representative deep learning model, the convolutional layer of the convolutional neural network (CNN) is the key to efficiently extracting input data features. The original EMG signal is a one-dimensional time series signal, and the constructed data set is a one-dimensional numerical data that reflects the time domain characteristics of EMG. The traditional CNN has few bottom-level convolution kernels and many high-level convolution kernels. This structure is prone to gradient explosion and overfitting when processing one-dimensional numerical data. Therefore, based on the characteristics of the one-dimensional raw data of the EMG signal, it is crucial to construct a suitable deep learning model to automatically extract signal features that are conducive to distinguishing CTS, which is crucial for realizing the intelligent diagnosis and classification of CTS. Summary of the invention

[0006] In view of the problems and shortcomings of the prior art, the present invention provides a carpal tunnel syndrome intelligent diagnosis system based on DPCNN.

[0007] The present invention solves the above technical problems through the following technical solutions:

[0008] The present invention provides a carpal tunnel syndrome intelligent diagnosis system based on DPCNN, which is characterized in that it includes an electromyographic data acquisition module, an electromyographic data preprocessing module, a data set construction module, a deep learning model construction module, a model training and testing module, a model cross-validation module, an optimal model output module and a model prediction and diagnosis module;

[0009] The electromyographic data acquisition module is used to acquire electromyographic data of all test items during the CTS examination of multiple subjects at a set acquisition frequency to obtain multiple electromyographic data of each subject;

[0010] The electromyographic data preprocessing module is used to perform preprocessing operations on each electromyographic data to remove power frequency noise and peak amplitude interference;

[0011] The data set construction module is used to splice and integrate various pre-processed electromyographic data of a subject into an electromyographic sample data, so that each subject corresponds to an electromyographic sample data, and each electromyographic sample data is labeled with a diagnostic label according to the professional diagnosis report of the clinician, and the diagnostic label is a CTS normal control group label and CTS labels of different severity. A model data set is constructed based on each electromyographic sample data and the corresponding diagnostic label, and the model data set includes a training set and a test set;

[0012] The deep learning model construction module is used to select the DPCNN structure to build a deep learning model according to the characteristics of the electromyographic sample data as one-dimensional numerical data. The deep learning model includes multiple layers of feature extraction layers, fully connected layers, Dropout layers and Softmax fully connected layers in sequence. Each feature extraction layer includes a convolution layer, a batch normalization layer and an activation function layer. The number of convolution kernels of the previous convolution layer in the hierarchical order is greater than the number of convolution kernels of the next convolution layer. The optimizer of each convolution layer uses the Adam optimizer. The initialization hyperparameters: learning rate and batch size determine the size of the sample size used in each iteration during training and the number of Epochs. The learning rate value range is 1*10 -5 Up to 1*10 -3 , the batch size is a power of 2;

[0013] The model training and testing module is used to input each electromyographic sample data in the training set into the constructed deep learning model. The deep learning model extracts the sample data features for iterative training, and continuously adjusts the hyperparameters in combination with the model training results until the optimal solution is obtained;

[0014] The model cross-validation module is used to perform multi-fold cross-validation training on the constructed deep learning model using the model data set to ensure that all sample data in the model data set have participated in the model training as training sets and test sets. According to the cross-validation results, a confusion matrix is ​​constructed to comprehensively evaluate the model classification performance;

[0015] The optimal model output module is used to output the deep learning model with the best performance in various indicators in the cross-validation as the optimal deep learning model;

[0016] The model prediction diagnosis module is used to collect multiple electromyographic data of a new subject, splice and integrate them into a set of electromyographic sample data to be predicted after preprocessing, and input the electromyographic sample data to be predicted into the optimal deep learning model to obtain the corresponding output intelligent diagnosis result.

[0017] The positive and progressive effects of the present invention are:

[0018] The present invention provides an accurate and efficient CTS intelligent diagnosis system to replace manual work, greatly reduces the workload of clinical neurologists, improves the efficiency of clinical CTS diagnosis and grading, and has great clinical application prospects.

[0019] The present invention innovatively uses the DPCNN architecture, effectively solving the problems of gradient explosion and overfitting that are prone to occur in traditional CNN on one-dimensional numerical data, and improving the robustness and generalization ability of the algorithm.

[0020] The system provided by the present invention is based on original electromyographic data and realizes automatic extraction of CTS-related features, further reducing human intervention and increasing the objectivity of intelligent diagnosis results.

[0021] The present invention provides an effective way for clinical intelligent diagnosis CTS, which improves the efficiency of clinical diagnosis, indirectly reduces the cost of neurological examinations for patients, and reduces the economic burden on patients' families. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 The structure block diagram of the carpal tunnel syndrome intelligent diagnosis system according to the preferred embodiment of the present invention.

[0023] Figure 2 This is a deep learning model architecture diagram of a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0025] like Figure 1 As shown, this embodiment provides an intelligent diagnosis system for carpal tunnel syndrome based on DPCNN, which includes an electromyography data acquisition module, an electromyography data preprocessing module, a data set construction module, a deep learning model construction module, a model training and testing module, a model cross-validation module, an optimal model output module and a model prediction and diagnosis module.

[0026] The electromyographic data acquisition module is used to acquire electromyographic data of all test items in the CTS examination process of multiple subjects at a set acquisition frequency to obtain multiple electromyographic data of each subject.

[0027] For example, at a sampling frequency of 1000 Hz, the EMG data of 100 subjects were collected 1 second after stimulation of each item in the clinical CTS test. The EMG data of all test items include: (1) EMG data of the median nerve from wrist to elbow collected from the abductor pollicis brevis in the motor conduction test; (2) EMG data of the ulnar nerve from wrist to 5 cm below the elbow collected from the abductor digiti minimi in the motor conduction test; (3) EMG data of the ulnar nerve from 5 cm below the elbow to 5 cm above the elbow collected from the abductor digiti minimi in the motor conduction test; (4) EMG data of the median nerve from middle finger to wrist collected from the middle finger in the sensory conduction test; (5) EMG data of the ulnar nerve from the little finger to wrist collected from the little finger in the sensory conduction test.

[0028] The EMG data preprocessing module is used to perform preprocessing operations on each EMG data to remove power frequency noise and peak amplitude interference, effectively retain the EMG features related to CTS in the collected original EMG data, and reduce the impact of noise artifacts on the model feature extraction performance and data quality.

[0029] For example: first perform a fast Fourier transform on each EMG data to convert the time series signal of the EMG data into a single-sided amplitude spectrum, then use a digital notch filter to filter each single-sided amplitude spectrum to remove the power frequency noise with large interference in the single-sided amplitude spectrum, and then use a Butterworth digital filter to remove the high-frequency peak amplitude interference in each single-sided amplitude spectrum after filtering.

[0030] The data set construction module is used to splice and integrate the pre-processed EMG data of a subject into an EMG sample data, so that each subject corresponds to an EMG sample data, and each EMG sample data is labeled with a diagnostic label according to the professional diagnosis report of the clinician. The diagnostic labels are CTS normal control group, mild CTS and severe CTS. A model data set is constructed based on each EMG sample data and the corresponding diagnostic label. The model data set includes a training set and a test set.

[0031] For example, each subject collects 1 second of EMG data for each of the five examination items, and each EMG data segment contains 1,000 time series data points. The 1 second EMG data collected for each examination item is merged into 5 seconds of electromyographic sample data, and each merged electromyographic sample data contains 5,000 time series data points. According to the clinician's diagnosis, each electromyographic sample data is marked with a diagnostic label of CTS control group, mild CTS, and severe CTS. Each electromyographic sample data corresponds to the diagnostic label one by one to construct a model data set.

[0032] The deep learning model building module is used to select the DPCNN structure to build a deep learning model based on the characteristics of the electromyographic sample data as one-dimensional numerical data. The deep learning model includes 4 layers of feature extraction layer, fully connected layer, Dropout layer and Softmax fully connected layer in sequence. Each feature extraction layer includes a convolution layer, a batch normalization layer and an activation function layer. The number of convolution kernels in the previous convolution layer of the multi-layer convolution layer sorted by level is greater than the number of convolution kernels in the next convolution layer. The optimizer of each convolution layer uses the Adam optimizer. The initialization hyperparameters: learning rate and batch size determine the size of the sample used in each iteration during training and the number of Epochs. The learning rate value range is 1*10 -5 Up to 1*10 -3 , the batch size is a power of 2.

[0033] The feature extraction layer of the deep learning model has 4 layers. In order to reduce the learnable parameters of the model and reduce the risk of overfitting, the convolution kernels in each convolution layer are designed to decrease as the number of layers increases. The first convolution layer contains 24 convolution kernels, the second convolution layer contains 16 convolution kernels, the third convolution layer contains 8 convolution kernels, and the fourth convolution layer contains 4 convolution kernels.

[0034] According to the characteristics of the EMG sample data set, a suitable deep learning framework is selected to build a deep learning model suitable for one-dimensional EMG data. This deep learning model uses the DPCNN architecture. The number of convolution kernels in each convolution layer decreases as the number of layers increases, effectively avoiding problems such as overfitting and gradient explosion that are prone to occur when training the model on one-dimensional numerical data. This solution can effectively avoid the gradient explosion problem that is prone to occur in traditional CNN while efficiently extracting EMG data features.

[0035] The model training and testing module is used to input the electromyographic sample data in the training set into the constructed deep learning model. The deep learning model extracts the sample data features for iterative training, and continuously adjusts the hyperparameters based on the model training results until the optimal solution is obtained.

[0036] Among them, the first convolution layer extracts the features of each input electromyographic sample data, and inputs the feature data into the first batch normalization layer for normalization processing. The second to fourth convolution layers extract the feature data input from the previous convolution layer, and input the feature data into the batch normalization layer of the corresponding layer for normalization processing. The normalized feature data is delinearized by the activation function and then passed to the next layer as the output data of the feature extraction layer. The output data of the previous feature extraction layer is used as the input data of the next layer. The input data completes the convolution and normalization processing in the corresponding convolution layer and batch normalization layer.

[0037] First, calculate the mean u of the hidden activation of the h-th layer batch input feature value, and the normalization formula is as follows:

[0038]

[0039] Among them, h i is a feature data point of the hth layer and the i-th batch, ∑h i is the total number of feature data points in the h-th layer and the i-th batch, m is the number of neurons in the h-th layer, h is the h-th layer in the number of feature extraction layers, and the diagnostic labels corresponding to each EMG sample data input in the i-th batch are all consistent.

[0040] Next, we calculate the standard deviation α of the hidden activations using the following formula:

[0041]

[0042] Then the input feature data of the hth layer is normalized to h i(norm) , the formula is as follows:

[0043]

[0044] Wherein, δ is a constant smoothing term that is not zero but close to zero, and δ may be 0.0000001 to prevent the denominator from being zero to ensure numerical stability in the operation.

[0045] Finally, the input is rescaled and offset to ensure that each output follows the standard normal distribution of the entire batch, as follows:

[0046] h i ′=γ*h i(norm) +β

[0047] Among them, h i ′ is the normalized feature data of the i-th batch of the h-th layer, γ is the rescaling parameter, and β is the offset parameter.

[0048] The fully connected layer multiplies the feature matrix vector composed of the feature data input from the last feature extraction layer by the preset weight matrix, adds a bias term, and outputs the second feature matrix vector through the activation function mapping of the feature results.

[0049] To further prevent the model from overfitting training and improve the generalization ability, a Dropout layer is added after the fully connected layer. The Dropout layer discards certain eigenvalues ​​with a certain probability p during training, and retains the third eigenvalue matrix vector formed and inputs it into the fully connected layer with a Softmax classifier. The eigenvalues ​​discarded in each training are completely different, so that each training results in a unique deep learning model, and finally the good weight parameters in each model are integrated into one model. In this embodiment, Dropout (p) = 0.5.

[0050] The fully connected layer with Softmax classifier classifies the feature data corresponding to each batch in the third feature matrix vector, maps the classification results of each batch output to [0,1], that is, converts them into corresponding probabilities, selects the classification result with the highest probability as the output, and the sum of the classification probabilities is 1.

[0051] The deep learning model is trained iteratively, and the hyperparameters are continuously adjusted according to the accuracy training results output in real time during the training process until the optimal solution is obtained. The deep learning model corresponding to the optimal solution is the trained deep learning model, and the performance of the trained deep learning model is evaluated using the test set.

[0052] The model cross-validation module is used to perform multi-fold cross-validation training on the constructed deep learning model using the model dataset, ensuring that all sample data in the model dataset has participated in model training as training sets and test sets. Based on the cross-validation results, a confusion matrix is ​​constructed to comprehensively evaluate the model classification performance.

[0053] The model data set was divided into a training set and a test set in a ratio of 8:2. A 5-fold cross-validation was performed on the deep learning model. The diagnostic label of each EMG sample data in the test set was used as the horizontal axis data, and the classification result of each EMG sample data in the test set by the deep learning model was used as the vertical axis data. A confusion matrix was constructed, and the accuracy, precision, recall, and F1-score indicators were calculated through the confusion matrix to comprehensively evaluate the model performance.

[0054]

[0055]

[0056]

[0057]

[0058] Among them, when the CTS control group is positive, and the mild CTS and severe CTS are negative: TP = True Positive, that is, when the sample is predicted to be positive, the sample is actually positive. For example, the CTS control group is predicted to be the CTS control group.

[0059] TN = True Negative, that is, when a sample is predicted to be negative, the sample is actually negative. For example, mild CTS is predicted to be mild CTS, and severe CTS is predicted to be severe CTS.

[0060] FP = False Positive, that is, when a sample is predicted to be positive, it is actually negative. For example, mild CTS and severe CTS are predicted to be CTS control groups.

[0061] FN = False Negative, that is, when a sample is predicted to be negative, it is actually positive. For example, the CTS control group is predicted to be mild CTS and severe CTS.

[0062] The optimal model output module is used to save the weight parameters and structure of the deep learning model with the best performance in various indicators in cross-validation as the optimal deep learning model, export the saved optimal deep learning model and encapsulate it into an API interface, and deploy the optimal deep learning model on the server after building the API server and configuring the environment.

[0063] For example: After the deep learning model is trained, use Torchscript to save and export the model. Then, use the Flask web framework to encapsulate the exported deep learning model into an API interface. Then, build an API server and configure the deep learning model environment to receive client requests. Finally, deploy the API server on the cloud server so that the client can call the API interface to implement deep learning model inference operations.

[0064] The model prediction diagnosis module is used to collect multiple electromyographic data of a new subject, splice and integrate them into a piece of electromyographic sample data to be predicted after preprocessing, upload the electromyographic sample data to be predicted to the server and call the API interface, so that the electromyographic sample data to be predicted is input into the optimal deep learning model to obtain the corresponding output intelligent diagnosis result.

[0065] For example: the client collects EMG data of a new subject 1 second after stimulation of the median nerve and ulnar nerve in the motor conduction measurement and sensory conduction measurement, and uploads it to the cloud server through the client after preprocessing and calls the API interface. The deep learning model performs calculations and reasoning on the uploaded data, and the output results are then transmitted back to the host client to complete the intelligent diagnosis of CTS.

[0066] In this solution, optimizer: The optimizer determines the update method of the weights of the deep learning model during the training process, and plays a decisive role in the performance of the deep learning model. Technicians have conducted comparative studies on the four common optimizers: SGD, Adagrad, PMSProp and Adam. Although the model using the SGD optimizer has a very fast training speed, the convergence speed of the model is very slow due to the large change in the parameter iteration direction. Adagrad is the earliest adaptive optimizer. The learning rate of the model using the Adagrad optimizer will monotonically decrease to 0, and the training process of the model is easy to be terminated early. RMSProp uses the exponential weighted average method to calculate the second-order momentum, which changes the problem of monotonically decreasing learning rate in Adagrad. Adam is a further improvement of RMSprop. Among the four models, the present invention finally selects the Adam optimizer.

[0067] In this solution, activation function: The activation function is the core of the neural network, which performs nonlinear transformation on the input to enable it to perform complex tasks. The present invention tested three activation functions: tanh, ReLU and Leaky ReLU, among which LeakyReLU has the best effect. The Leaky ReLU activation function has high computational efficiency and helps the deep learning model converge quickly. When the input value is negative, the function still has a small gradient, which solves the gradient vanishing problem when using the ReLU activation function. Therefore, the deep learning model finally chooses the Leaky ReLU activation function.

[0068] In this solution, learning rate: The learning rate determines the learning efficiency of the model weights and is one of the most influential parameters. -5 Up to 1*10 -3 We conducted multiple tests and finally the performance of the model was set to 2*10 -5 The best time.

[0069] In this solution, batch size: batch size determines the size of the sample size used in each iteration of the training process. A batch size that is too small will lead to reduced model training efficiency and failure to converge within a certain number of Epochs. A batch size that is too large will lead to poor generalization ability of the model, and the trained model will perform poorly on the test set. The present invention compares the performance of models with batch sizes of 8, 16, 32, and 64. Among them, 8 and 16 performed well. Considering the operating efficiency of the model, we finally chose a batch size of 16.

[0070] The parameter optimization of this scheme first improves the updating method of model weights by selecting a suitable optimizer; secondly, it selects a suitable activation function according to the data type of the data sample to improve the model training speed, accelerate the model convergence and avoid gradient vanishing; then, it adjusts the model weight learning rate in real time according to the training effect of the model; finally, it selects a suitable batch size to ensure that the sample batch size of the input model is appropriate, so that the model converges quickly on the training set and has high generalization ability and robustness on the test set.

[0071] After data collection, data set construction, deep learning model construction, and model hyperparameter optimization are completed, the model can be iteratively trained. Epoch represents the process in which all training sample data is propagated once in the model in the forward and reverse directions. It is the basic unit of iterative model training. In order to better observe the training effect of the model, we observe the training loss and validation loss of each epoch and select the appropriate number of epochs to prevent overfitting when the model loss converges or underfitting when the loss does not converge.

[0072] Cross-validation first divides the sample data set into a training set and a test set. There are many combinations of the division of the training set and the test set. In each combination, the training set is used to train the model, and the validation set is used to evaluate the performance of the model. Since the combination of the data set after each division is different, after training and validating the data of this group, the training set data may be used as the test set data, and the test set data may be used as the training set data in the next training and validation process. Cross-validation is to reuse sample data so that each data sample has been used as a training set and a test set. Cross-validation can help the model effectively avoid problems caused by unreasonable data set division, making the model performance evaluation more comprehensive and objective.

[0073] The confusion matrix displays the relationship between the model's classification results on the test set and the true labels in the form of a matrix. In addition to showing the number of correctly classified samples, it can also reveal the specific misclassification patterns of the model.

[0074] Based on cross-validation and confusion matrix, the performance evaluation provides multiple performance evaluation indicators such as classification accuracy, precision, recall rate, and F1-score of the deep learning model to achieve a comprehensive and all-round evaluation of the model performance.

[0075] Although the specific embodiments of the present invention are described above, those skilled in the art should understand that these are only examples, and the protection scope of the present invention is defined by the appended claims. Those skilled in the art may make various changes or modifications to these embodiments without departing from the principles and essence of the present invention, but these changes and modifications all fall within the protection scope of the present invention.

Claims

1. An intelligent diagnosis system for carpal tunnel syndrome based on DPCNN, characterized in that: It includes an electromyographic data acquisition module, an electromyographic data preprocessing module, a data set construction module, a deep learning model construction module, a model training and testing module, a model cross-validation module, an optimal model output module and a model prediction and diagnosis module; The electromyographic data acquisition module is used to acquire electromyographic data of all test items during the CTS examination of multiple subjects at a set acquisition frequency to obtain multiple electromyographic data of each subject; The electromyographic data preprocessing module is used to perform preprocessing operations on each electromyographic data to remove power frequency noise and peak amplitude interference; The data set construction module is used to splice and integrate various pre-processed electromyographic data of a subject into an electromyographic sample data, so that each subject corresponds to an electromyographic sample data, and each electromyographic sample data is labeled with a diagnostic label according to the professional diagnosis report of the clinician, and the diagnostic label is a CTS normal control group label and CTS labels of different severity. A model data set is constructed based on each electromyographic sample data and the corresponding diagnostic label, and the model data set includes a training set and a test set; The deep learning model construction module is used to select the DPCNN structure to build a deep learning model according to the characteristics of the electromyographic sample data as one-dimensional numerical data. The deep learning model includes multiple layers of feature extraction layers, fully connected layers, Dropout layers and Softmax fully connected layers in sequence. Each feature extraction layer includes a convolution layer, a batch normalization layer and an activation function layer. The number of convolution kernels of the previous convolution layer in the hierarchical order is greater than the number of convolution kernels of the next convolution layer. The optimizer of each convolution layer uses the Adam optimizer. The initialization hyperparameters: learning rate and batch size determine the size of the sample size used in each iteration during training and the number of Epochs. The learning rate value range is 1*10 -5 Up to 1*10 -3 , the batch size is a power of 2; The model training and testing module is used to input each electromyographic sample data in the training set into the constructed deep learning model. The deep learning model extracts the sample data features for iterative training, and continuously adjusts the hyperparameters in combination with the model training results until the optimal solution is obtained; The model cross-validation module is used to perform multi-fold cross-validation training on the constructed deep learning model using the model data set to ensure that all sample data in the model data set have participated in model training as training sets and test sets. According to the cross-validation results, a confusion matrix is ​​constructed to comprehensively evaluate the model classification performance; The optimal model output module is used to output the deep learning model with the best performance in various indicators in the cross-validation as the optimal deep learning model; The model prediction diagnosis module is used to collect multiple electromyographic data of a new subject, splice and integrate them into a piece of electromyographic sample data to be predicted after preprocessing, and input the electromyographic sample data to be predicted into the optimal deep learning model to obtain the corresponding output intelligent diagnosis result; The model training and testing module is used to: The first convolution layer extracts the features of each input electromyographic sample data, and inputs the feature data into the first batch normalization layer for normalization. The normalized feature data is delinearized by the activation function and then passed to the next layer as the output data of the feature extraction layer. The output data of the previous feature extraction layer is used as the input data of the next layer. The input data completes convolution and normalization processing in the corresponding convolution layer and batch normalization layer. First calculate the h Mean of the hidden activations of the layer batch input feature values u , the normalization formula is as follows: in, h i It is h Tier i A characteristic data point of a batch, It is h Tier i The total number of feature data points in the batch, m It is h The number of neurons in the layer, h It is the number of feature extraction layers. h Layer, i The diagnostic labels corresponding to each EMG sample data input in the batch are consistent; Next, we calculate the standard deviation of the hidden activations α , the formula is as follows: Then the h The input feature data of the layer is normalized to , the formula is as follows: in, δ is a constant smoothing term that is not zero but close to zero; Finally, the input is rescaled and offset to ensure that each output follows the standard normal distribution of the entire batch, as follows: in, For the h Tier i The normalized feature data of the batch, γ is the rescaling parameter, β is the offset parameter; The fully connected layer multiplies the feature matrix vector composed of the feature data input from the last feature extraction layer by the preset weight matrix, adds a bias term, and outputs the second feature matrix vector through the activation function mapping of the feature result; The Dropout layer is trained with a certain probability p Some eigenvalues ​​in the second feature matrix vector are discarded, and the third feature matrix vector formed by retaining the remaining eigenvalues ​​is input into the fully connected layer with the Softmax classifier; The fully connected layer with Softmax classifier classifies the feature data corresponding to each batch in the third feature matrix vector, maps the classification results of each batch output to [0,1], that is, converts them into corresponding probabilities, selects the classification result with the highest probability as the output, and the sum of the classification probabilities is 1; The deep learning model is iteratively trained. Combined with the accuracy training results output in real time during the training process, the hyperparameters are continuously adjusted until the optimal solution is obtained. The deep learning model corresponding to the optimal solution is the trained deep learning model, and the performance of the trained deep learning model is evaluated using the test set. The diagnostic labels were CTS normal control group, mild CTS, and severe CTS, and the model dataset was divided into training and testing sets in a ratio of 8:2; The model cross-validation module is used to perform 5-fold cross-validation on the deep learning model, taking the diagnostic label of each electromyographic sample data in the test set as the horizontal axis data, and the classification result of each electromyographic sample data in the test set by the deep learning model as the vertical axis data, to construct a confusion matrix, and calculate the accuracy, precision, recall, and F1-score indicators through the confusion matrix to comprehensively evaluate the model performance; The feature extraction layer has 4 layers, the first convolution layer contains 24 convolution kernels, the second convolution layer contains 16 convolution kernels, the third convolution layer contains 8 convolution kernels, and the fourth convolution layer contains 4 convolution kernels; The electromyographic data of all test items include: (1) the EMG data of the median nerve from the wrist to the elbow collected from the abductor pollicis brevis in the motor conduction measurement; (2) the EMG data of the ulnar nerve from the wrist to 5 cm below the elbow collected from the abductor digiti minimi in the motor conduction measurement; (3) the EMG data of the ulnar nerve from 5 cm below the elbow to 5 cm above the elbow collected from the abductor digiti minimi in the motor conduction measurement; (4) the EMG data of the median nerve from the middle finger to the wrist collected from the middle finger in the sensory conduction measurement; (5) the EMG data of the ulnar nerve from the little finger to the wrist collected from the little finger in the sensory conduction measurement.

2. The DPCNN-based intelligent diagnosis system for carpal tunnel syndrome according to claim 1, characterized in that: The electromyographic data preprocessing module is used to first perform a fast Fourier transform on each electromyographic data, convert the time series signal of the electromyographic data into a single-sided amplitude spectrum, and then use a digital notch filter to filter each single-sided amplitude spectrum to remove the power frequency noise with large interference in the single-sided amplitude spectrum, and then use a Butterworth digital filter to remove the high-frequency peak amplitude interference in each single-sided amplitude spectrum after filtering.

3. The DPCNN-based intelligent diagnosis system for carpal tunnel syndrome according to claim 1, characterized in that: The optimal model output module is used to save the weight parameters and structure of the deep learning model with the best performance in various indicators in the cross-validation as the optimal deep learning model, export the saved optimal deep learning model and encapsulate it into an API interface, and deploy the optimal deep learning model on the server after the API server is built and the environment is configured; The model prediction and diagnosis module is used to collect multiple electromyographic data of a new subject, splice and integrate them into a piece of electromyographic sample data to be predicted after preprocessing, upload the electromyographic sample data to be predicted to the server and call the API interface, so that the electromyographic sample data to be predicted is input into the optimal deep learning model to obtain the corresponding output intelligent diagnosis result.

4. The DPCNN-based intelligent diagnosis system for carpal tunnel syndrome according to claim 1, characterized in that: In the optimal deep learning model, the learning rate is 2*10 -5 , the batch size is 16.

Images