Ventricular premature beat recognition method based on improved MobileNetV3

Through the improved MobileNetV3 model, combined with bottleneck module and jump module, the problem of complex memory occupancy of deep convolutional neural network model is solved, and efficient ventricular premature beat recognition on mobile devices is achieved, with a classification accuracy of more than 90%.

CN114886435BInactive Publication Date: 2024-11-22GUANGXI NORMAL UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210507171.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-05-10
Publication Date
2024-11-22
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The existing deep convolutional neural network models are complex and occupy a large amount of memory resources, resulting in slow training speed and difficult to meet the requirements of mobile devices.

Method used

Using the improved MobileNetV3 model, the ventricular premature beat recognition model consisting of 3 convolutional layers, 5 bottleneck modules, 2 jump modules, 2 average pooling layers and 1 global average pooling layer are constructed, combining depth The convolutional layer and the compression-exciting attention mechanism layer can be separated, reducing the amount of parameters and improving network real-time.

Benefits of technology

A ventricular premature beat recognition model with small parameters and low delay is realized. It is suitable for deployment on intelligent mobile devices, with a classification accuracy of more than 90%, meeting the real-time requirements of mobile devices.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114886435B_ABST
    Figure CN114886435B_ABST
Patent Text Reader

Abstract

The present invention discloses a ventricular premature beat recognition method based on an improved MobileNetV3, and proposes an improved MobileNetV3 model. First, the constructed ventricular premature beat recognition model is trained using training samples, and then the predicted samples are sent to the trained ventricular premature beat recognition model to achieve the classification of the electrocardiogram to be classified. The various indicators obtained by the MobileNetV3 network with Skipblock are higher than those of the MobileNetV3 network without Skipblock; the MobileNetV3 network model has the characteristics of small number of parameters and low latency, which solves the problem that wearable devices such as watches and bracelets generally have high requirements on model parameters and network real-time performance, and the models with large number of parameters and complex structures are not suitable for deployment on smart mobile terminals.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of premature heart beat recognition, and in particular to a premature ventricular contraction (PVC) recognition method based on an improved MobileNetV3. Background Art

[0002] Arrhythmia is a common cardiovascular disease in clinic, and premature beats are the most common arrhythmia. By introducing computer-assisted premature beat analysis technology, doctors can carry out effective medical intervention as early as possible, which plays a very important role in the prevention and treatment of diseases. It can also solve the problem of scarce medical resources in remote mountainous areas and the inability to enjoy basic medical services to a certain extent.

[0003] Currently, deep convolutional neural networks have achieved relatively high accuracy in multiple classification tasks. However, deep convolutional neural networks have a large number of parameters, require a large amount of memory for the model, take a long time to train, and require computer hardware to have greater computing power. Summary of the invention

[0004] The present invention aims to solve the problem that the existing deep convolutional neural network model is complex and occupies a large amount of memory resources, making it difficult to meet the requirements of mobile devices in terms of training speed, and provides a ventricular premature beat recognition method based on an improved MobileNetV3.

[0005] To solve the above problems, the present invention is achieved through the following technical solutions:

[0006] The ventricular premature beat recognition method based on the improved MobileNetV3 includes the following steps:

[0007] Step 1: construct a ventricular premature beat recognition model;

[0008] The above-mentioned ventricular premature beat recognition model consists of 3 convolutional layers, 1 maximum pooling layer, 5 bottleneck modules, 2 skip connection modules, 1 average pooling layer, 1 average pooling layer, 1 global average pooling layer and 1 output layer; the input of the first convolutional layer is used as the input of the ventricular premature beat recognition model; the input of the first convolutional layer is connected to the input of the maximum pooling layer, and the output of the maximum pooling layer is connected to the input of the first bottleneck module and the two skip connection modules; the output of the first bottleneck module is connected to the input of the second bottleneck module, the output of the second bottleneck module is connected to the input of the third bottleneck module, the output of the first bottleneck module is added to the output of the first skip connection module and then connected to the input of the fourth bottleneck module, the output of the fourth bottleneck module is connected to the input of the fifth bottleneck module, and the output of the fifth bottleneck module is added to the output of the second skip connection module and then connected to the input of the second convolutional layer; the output of the second convolutional layer is connected to the input of the average pooling layer, the output of the average pooling layer is connected to the input of the third convolutional layer, and the output of the third convolutional layer is connected to the input of the first global average pooling layer; the output of the first global average pooling layer is connected to the input of the output layer, and the output of the output layer is used as the output of the ventricular premature beat recognition model;

[0009] Step 2: pre-process the labeled and classified electrocardiograms as training samples, and use the training samples to train the ventricular premature beat recognition model constructed in step 1 to obtain a trained ventricular premature beat recognition model;

[0010] Step 3: Preprocess the electrocardiogram to be classified as a prediction sample, and send the prediction sample to the ventricular premature beat recognition model trained in step 2 to classify the electrocardiogram to be classified.

[0011] In the above scheme, the bottleneck module consists of 2 convolutional layers, 1 depth-wise separable convolutional layer and a compression-excitation attention mechanism layer; the input of the fourth convolutional layer is used as the input of the bottleneck module; the output of the fourth convolutional layer is connected to the input of the first depth-wise separable convolutional layer after batch normalization BN and non-linear activation function NL; the output of the first depth-wise separable convolutional layer is connected to the input of the compression-excitation attention mechanism layer after batch normalization BN and non-linear activation function NL; the output of the compression-excitation attention mechanism layer is connected to the input of the fifth convolutional layer; the output of the fifth convolutional layer is added to the input of the fourth convolutional layer after batch normalization BN and used as the output of the bottleneck module.

[0012] In the above scheme, the compression-excitation attention mechanism layer consists of 1 global average pooling layer, 2 fully connected layers, 1 ReLU activation function layer and 1 h-sigmoid activation function layer; the input of the second global average pooling layer is used as the input of the compression-excitation attention mechanism layer; the output of the second global average pooling layer is connected to the input of the first fully connected layer, the output of the first fully connected layer is connected to the input of the ReLU activation function layer, the output of the ReLU activation function layer is connected to the input of the second fully connected layer, the output of the second fully connected layer is connected to the input of the h-sigmoid activation function layer, and the output of the h-sigmoid activation function layer is multiplied by the input of the second global average pooling layer as the output of the compression-excitation attention mechanism layer.

[0013] In the above scheme, the skip connection module consists of 2 convolutional layers and 1 depth-wise separable convolutional layer; the input of the sixth convolutional layer is used as the input of the skip connection module; the output of the sixth convolutional layer is connected to the input of the second depth-wise separable convolutional layer, the output of the second depth-wise separable convolutional layer is connected to the input of the seventh convolutional layer, and the output of the seventh convolutional layer is used as the output of the skip connection module.

[0014] Compared with the prior art, the present invention has the following characteristics:

[0015] 1. The MobileNetV3 network model has the characteristics of small number of parameters and low latency, which solves the problem that wearable devices such as watches and bracelets generally have high requirements for model parameters and network real-time performance, so models with large number of parameters and complex structure are not suitable for deployment on smart mobile terminals.

[0016] 2. The idea of ​​the Skipblock module is to directly transfer low-level features from the first layer to the last layer, and use the low-level features extracted from the first layer to enrich the last layer. Through experimental comparison, it is found that the various indicators obtained by the MobileNetV3 network with Skipblock are higher than those without Skipblock. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is a structural diagram of the ventricular premature beat recognition model MobileNetV3.

[0018] Figure 2 for Figure 1 Schematic diagram of the structure of the bottleneck module Bneck.

[0019] Figure 3 for Figure 2 Schematic diagram of the structure of the SE attention mechanism layer.

[0020] Figure 4 for Figure 1 Schematic diagram of the structure of the mid-skip connection module Skipbolck. DETAILED DESCRIPTION

[0021] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific examples.

[0022] A ventricular premature beat recognition method based on improved MobileNetV3 includes the following steps:

[0023] Step 1: Construct a ventricular premature beat recognition model.

[0024] The MobileNet model is a lightweight deep neural network proposed by Google in recent years for use in mobile devices and embedded platforms. Compared with traditional deep learning models, the MobileNet model has the characteristics of fewer parameters and low latency. The present invention improves the MobileNetV3 model, proposes a ventricular premature beat recognition model MobileNetV3_Small based on the improved MobileNetV3 model, and uses it for ventricular premature beat recognition.

[0025] The above ventricular premature beat recognition model MobileNetV3_Small is as follows Figure 1 As shown in FIG. 1 , it consists of 3 convolutional layers (Conv), 1 maximum pooling layer (Max Pool), 5 bottleneck modules (Bneck), 2 skip modules (Skipbolck), 1 average pooling layer (Average Pool), 1 average pooling layer (Average Pool), 1 global average pooling layer (Global Pool) and 1 output layer. The input of the first convolutional layer is used as the input of the ventricular premature beat recognition model. The input of the first convolutional layer is connected to the input of the maximum pooling layer, and the output of the maximum pooling layer is connected to the input of the first bottleneck module and the two skip modules. The output of the first bottleneck module is connected to the input of the second bottleneck module, the output of the second bottleneck module is connected to the input of the third bottleneck module, the output of the first bottleneck module is added to the output of the first skip module, and the input of the fourth bottleneck module is connected to the input of the fifth bottleneck module, and the output of the fifth bottleneck module is added to the output of the second skip module and connected to the input of the second convolutional layer. The output of the second convolutional layer is connected to the input of the average pooling layer, the output of the average pooling layer is connected to the input of the third convolutional layer, the output of the third convolutional layer is connected to the input of the first global average pooling layer, and the output of the first global average pooling layer is connected to the input of the output layer. The output of the output layer is used as the output of the ventricular premature beat recognition model.

[0026] The bottleneck module in the ventricular premature beat recognition model is as follows Figure 2As shown in the figure, it consists of 2 convolutional layers (Conv), 1 depthwise separable convolutional layer (DWConv) and a squeeze-excitation (SE) attention mechanism layer. The input of the fourth convolutional layer is used as the input of the bottleneck module. The output of the fourth convolutional layer is connected to the input of the first depthwise separable convolutional layer after batch normalization BN and nonlinear activation function NL. The output of the first depthwise separable convolutional layer is connected to the input of the squeeze-excitation attention mechanism layer after batch normalization BN and nonlinear activation function NL. The output of the squeeze-excitation attention mechanism layer is connected to the input of the fifth convolutional layer. The output of the fifth convolutional layer is added to the input of the fourth convolutional layer after batch normalization BN and used as the output of the bottleneck module.

[0027] The compression-excitation attention mechanism layer is as follows Figure 3 As shown in the figure, it consists of 1 global average pooling layer (Global Pool), 2 fully connected layers (FC), 1 ReLU activation function layer and 1 h-sigmoid activation function layer. The input of the second global average pooling layer is used as the input of the compression-excitation attention mechanism layer. The output of the second global average pooling layer is connected to the input of the first fully connected layer, the output of the first fully connected layer is connected to the input of the ReLU activation function layer, the output of the ReLU activation function layer is connected to the input of the second fully connected layer, the output of the second fully connected layer is connected to the input of the h-sigmoid activation function layer, and the output of the h-sigmoid activation function layer is multiplied with the input of the second global average pooling layer as the output of the compression-excitation attention mechanism layer. The SE attention module includes 1 global average pooling layer and 2 fully connected layers. The activation function of the first fully connected layer is ReLU, and the activation function of the second fully connected layer is h-sigmiod. The SE attention module automatically obtains the importance G between each feature channel by learning, and enhances the weight of useful features and suppresses features that have little effect on the current task according to the importance. In the compression-excitation attention mechanism layer, the input vector x passes through the SE attention module to obtain the weight value G of each feature map in the input vector x, and then the weight value G is multiplied by the input vector x to obtain the output vector.

[0028] In order to enrich the features of the high-level layers in the model, the present invention introduces the skipping module into the ventricular premature beat recognition model. The idea of ​​the skipping module is to directly transfer the low-level features from the bottom layer to the high-level layers. Figure 4 As shown in the figure, it consists of 2 convolutional layers and 1 depth-wise separable convolutional layer (DWConv). The input of the sixth convolutional layer is used as the input of the skip connection module. The output of the sixth convolutional layer is connected to the input of the second depth-wise separable convolutional layer, and the output of the second depth-wise separable convolutional layer is connected to the input of the seventh convolutional layer. The output of the seventh convolutional layer is used as the output of the skip connection module.

[0029] Step 2: Preprocess the labeled and classified electrocardiograms as training samples, and use the training samples to train the ventricular premature beat recognition model constructed in step 1 to obtain a trained ventricular premature beat recognition model.

[0030] The preprocessing process of ECG mainly uses filters to remove noise such as baseline drift and power frequency interference in ECG. The classification marked in the electrocardiogram is premature beat or non-premature beat.

[0031] Step 3: Preprocess the electrocardiogram to be classified as a prediction sample, and send the prediction sample to the ventricular premature beat recognition model trained in step 2 to classify the electrocardiogram to be classified.

[0032] The preprocessing method for the ECG to be classified is the same as the preprocessing method for the ECG with the labeled classification. The output of the prediction sample after the model is the premature beat or non-premature beat class record.

[0033] The performance of the present invention is described below through a specific example:

[0034] The parameters of each layer of the ventricular premature beat recognition model MobileNetV3 in this embodiment are shown in Table 1, where the Operator column represents the feature layer used and its convolution kernel or pooling kernel size, the exp size column is the number of feature maps of the bottleneck module dimension expansion layer, i.e., the first 1*1 convolution layer, #out is the number of feature maps of the last 1*1 convolution layer of the bottleneck module, SE indicates whether the bottleneck module introduces the SE attention mechanism, NL indicates the type of activation function, HS represents h-swish, RE represents ReLU, and stride is the step size of the convolution kernel or the pooling kernel.

[0035] Table 1 MobileNetV3 model in PVC recognition

[0036]

[0037] The Chinese Cardiovascular Disease Database (CCDD) was used as the basis, and the ECG records in the database were first denoised by a 0.5-40 Hz bandpass filter.

[0038] Since the premature beat classification of the present invention is a binary classification problem (premature beat or non-premature beat class), sensitivity (Se), specificity (Sp), and accuracy (Acc) can be used to measure the quality of the classification effect. The confusion matrix of the binary classification is shown in Table 2 below:

[0039] Table 2 Confusion matrix

[0040]

[0041] The definitions of each indicator are as follows:

[0042] Sensitivity (Se):

[0043] Se=TP / (TP+FN)

[0044] Specificity (Sp):

[0045] Sp=TN / (TN+FP)

[0046] Accuracy (Acc):

[0047] Acc=(TP+TN) / (TP+TN+FP+FN)

[0048] 35840 preprocessed ECG records were used as training samples, of which 3112 were ventricular premature beats (PVCs); 141046 records (including 2148 PVC records) were used as test samples. After the improved MobileNetV3 model was trained using the training samples, the results obtained by the model on the 141046 records, i.e., the test samples, are shown in Table 3 below, where NPVC represents non-ventricular premature beat records, PVC represents ventricular premature beat records, Se represents sensitivity, Sp represents specificity, and Acc represents accuracy.

[0049] Table 3 PVC recognition results of the improved MobileNetV3 model

[0050]

[0051] It can be seen from Table 3 that the sensitivity, specificity and accuracy of electrocardiogram classification using the ventricular premature beat recognition model MobileNetV3_Small of the present invention all reached more than 90%.

[0052] It should be noted that although the embodiments of the present invention described above are illustrative, they are not intended to limit the present invention, and therefore the present invention is not limited to the above specific embodiments. Without departing from the principles of the present invention, any other embodiments obtained by those skilled in the art under the guidance of the present invention are deemed to be within the protection of the present invention.

Claims

1. Ventricular premature beat recognition method based on improved MobileNetV3, Its characteristics are: The steps include: Step 1: construct a ventricular premature beat recognition model; The above ventricular premature beat recognition model consists of 3 convolutional layers, 1 maximum pooling layer, 5 bottleneck modules, 2 skip connection modules, 1 average pooling layer, 1 average pooling layer, 1 global average pooling layer and 1 output layer; The skip connection module consists of 2 convolutional layers and 1 depth-wise separable convolutional layer; the input of the sixth convolutional layer is used as the input of the skip connection module; the output of the sixth convolutional layer is connected to the input of the second depth-wise separable convolutional layer, the output of the second depth-wise separable convolutional layer is connected to the input of the seventh convolutional layer, and the output of the seventh convolutional layer is used as the output of the skip connection module; The input of the first convolutional layer is used as the input of the ventricular premature beat recognition model; the input of the first convolutional layer is connected to the input of the maximum pooling layer, and the output of the maximum pooling layer is connected to the input of the first bottleneck module and the two skip connection modules; the output of the first bottleneck module is connected to the input of the second bottleneck module, the output of the second bottleneck module is connected to the input of the third bottleneck module, the output of the first bottleneck module is added to the output of the first skip connection module and then connected to the input of the fourth bottleneck module, the output of the fourth bottleneck module is connected to the input of the fifth bottleneck module, the output of the fifth bottleneck module is added to the output of the second skip connection module and then connected to the input of the second convolutional layer; the output of the second convolutional layer is connected to the input of the average pooling layer, the output of the average pooling layer is connected to the input of the third convolutional layer, and the output of the third convolutional layer is connected to the input of the first global average pooling layer; the output of the first global average pooling layer is connected to the input of the output layer, and the output of the output layer is used as the output of the ventricular premature beat recognition model; Step 2: pre-process the labeled and classified electrocardiograms as training samples, and use the training samples to train the ventricular premature beat recognition model constructed in step 1 to obtain a trained ventricular premature beat recognition model; Step 3: Preprocess the electrocardiogram to be classified as a prediction sample, and send the prediction sample to the ventricular premature beat recognition model trained in step 2 to classify the electrocardiogram to be classified.

2. According to the method for identifying premature ventricular beats based on improved MobileNetV3 according to claim 1, Its characteristics are: The bottleneck module consists of 2 convolutional layers, 1 depth-wise separable convolutional layer and a compression-excitation attention mechanism layer; the input of the fourth convolutional layer is used as the input of the bottleneck module; the output of the fourth convolutional layer is connected to the input of the first depth-wise separable convolutional layer after batch normalization BN and nonlinear activation function NL; The output of the first depth-separable convolutional layer is batch normalized (BN) and the nonlinear activation function (NL) and then connected to the input of the compression-excitation attention mechanism layer; The output of the compression-excitation attention mechanism layer is connected to the input of the fifth convolutional layer; the output of the fifth convolutional layer is added to the input of the fourth convolutional layer after batch normalization (BN) and serves as the output of the bottleneck module.

3. According to the method for identifying premature ventricular beats based on improved MobileNetV3 according to claim 2, Its characteristics are: The compression-excitation attention mechanism layer consists of 1 global average pooling layer, 2 fully connected layers, 1 ReLU activation function layer, and 1 h-sigmoid activation function layer; The input of the second global average pooling layer is used as the input of the squeeze-excite attention mechanism layer; The output of the second global average pooling layer is connected to the input of the first fully connected layer, the output of the first fully connected layer is connected to the input of the ReLU activation function layer, the output of the ReLU activation function layer is connected to the input of the second fully connected layer, the output of the second fully connected layer is connected to the input of the h-sigmoid activation function layer, and the output of the h-sigmoid activation function layer is multiplied with the input of the second global average pooling layer as the output of the compression-excitation attention mechanism layer.