WO2021114761A1

WO2021114761A1 - Lung rale artificial intelligence real-time classification method, system and device of electronic stethoscope, and readable storage medium

Info

Publication number: WO2021114761A1
Application number: PCT/CN2020/113511
Authority: WO
Inventors: 蔡盛盛; 胡南; 刘仁雨; 徐兴国
Original assignee: 苏州美糯爱医疗科技有限公司
Priority date: 2019-12-13
Filing date: 2020-09-04
Publication date: 2021-06-17
Also published as: CN110970042A; CN110970042B

Abstract

A lung rale artificial intelligence real-time classification method of an electronic stethoscope. The method comprises: acquire a lung sound signal in real time by means of an electronic stethoscope, and automatically classify lung rale; slidingly extract the acquired data every two seconds as a data block, and normalize the data blocks by means of a band-pass filter; use a logarithm Mel filter bank transformation to calculate data matrixes of three channels, and input same into a pre-built and trained convolutional neural network, an output of the convolutional neural network being the probability values of four lung sound conditions; and a system gives the final probability values of the four lung sound conditions in combination with results of multiple data blocks. The robustness of a rale detection and classification result can be effectively improved. The present invention further provides a lung rale artificial intelligence real-time classification system and device of the electronic stethoscope and a computer readable storage medium, and the system and device have the same beneficial effects.

Description

Method, system, device and readable storage medium for artificial intelligence real-time classification of lung rales of electronic stethoscope

Technical field

The invention relates to the technical field of computer hearing and artificial intelligence, in particular to a method, system, device and readable storage medium for real-time artificial intelligence classification of pulmonary rales of an electronic stethoscope.

Background technique

Due to environmental pollution and deteriorating air quality, the incidence of various respiratory diseases such as asthma, pneumonia, and bronchitis is increasing year by year. Every year, nearly 1 million children under the age of 5 die from acute lower respiratory infections, and the number of deaths exceeds that of HIV, The sum of malaria and tuberculosis ulcers. As the disease of the respiratory system has become one of the diseases that seriously threaten human health, accurate diagnosis and effective treatment of respiratory diseases are effective ways to ensure the patient's recovery as soon as possible.

The methods currently used by hospitals to examine and identify respiratory diseases include: (1) Chest X-ray: This method can record general lung lesions, such as lung inflammation, masses, and tuberculosis. (2) Lung CT: This method helps to make a qualitative diagnosis of the problems found on chest X-rays, such as the type and location of the mass. (3) Bronchoscopy: This method is used to diagnose most lung and airway diseases. However, these methods are not only expensive but also relatively affect the human body, and due to geographical restrictions, some people may not be able to access these diagnostic methods.

Auscultation is one of the earliest and most direct examination methods for respiratory diseases. Medical staff mainly use stethoscopes to listen to whether the patient's breath sounds contain rales-mainly including wet rales and stridor. Figure 1 shows breath sounds with different additional rales, among which (a) contains wet rales, (b ) Contains wheezing sounds, (c) contains both rales and wheezing sounds, and (d) is a normal breathing sound. However, this method has always been restricted by factors such as the auscultation environment and the level of medical skills.

In the prior art, such as the digital stethoscope disclosed in Publication No. CN106022258A and the method for extracting lung sounds by filtering out heart sounds, discrete entropy values are used to first filter out part of the effective frames, and then the average amplitude of the filtered effective frames is extracted as the threshold. The threshold is used to obtain the lung sound frame containing the heart sound, and then the wavelet transform is performed and the threshold is used to filter out the relevant wavelet coefficients to obtain a relatively pure lung sound frame. The MFCC feature parameter matrix is extracted from the lung sound frame, and the feature parameter matrix is sent to the traditional Backward Propagation (BP) network for category judgment. This method needs to pass two threshold judgments, and relevant useful information will be lost in the threshold judgment, thereby reducing the effectiveness of the MFCC feature parameter matrix.

For example, CN107704885A discloses the method of realizing heart sound and lung sound classification on the intelligent platform. First, the received data is resampled at 5 points at a sampling frequency of 2205Hz. After the resampled signal is obtained, filter processing is performed, and the maximum bandpass attenuation is set to 3db , The minimum attenuation of the stop band is 18db. Then use dmey wavelet to perform wavelet denoising, and use autocorrelation coefficient to segment after the denoising signal is obtained. Then extract the MFCC feature parameter matrix for each segment and input the feature parameter matrix into a support vector machine (SVM) classifier for classification processing. However, the SVM classifier is not very efficient when processing high-dimensional data such as the MFCC feature parameter matrix, and this method does not provide a method that can be classified in real time.

For example, the paper "Pattern recognition methods applied to respiratory sounds classification into normal and wheeze classes" published by B. Mohammed combines MFCC features and Gaussian Mixture Model (GMM) to classify normal lung sounds and lung sounds with wheeze; P.Mayorga The published paper "Acoustics based assessment of respiratory disease using GMM classification" also uses GMM to classify the rales of lung sounds; the paper published by S. Alsmadi et al. "Design of a DSP-based instrument for real-time classification of "Pulmonary sounds" uses K-nearest neighbor (K-NN) and minimum distance criteria to judge whether the overall lung sounds are abnormal.

The method proposed in the above papers can classify a certain kind of rales or the overall situation of the data, but cannot make a comprehensive judgment on wet rales, wheezing, and multiple situations that contain both and none of them.

Another example is a lung sound classification method, system and application based on a convolutional neural network disclosed in the publication number CN107818366A. First, the lung sound signal is band-pass filtered, and then the lung sound timing signal is converted into two through short-time Fourier transform. Dimensional spectrogram, and finally use this spectrogram as an input feature to classify this lung sound signal. This patent is only a simple application of a convolutional neural network, and a simple binary conclusion of normal/abnormal lung sounds is obtained after inputting a signal of a fixed length. This method cannot meet real-time performance, is easily affected by short-time interference and causes misjudgment, and the classification result is too simple.

At present, the existing technologies for classifying lung rale signals mainly focus on traditional machine learning and pattern recognition. There are also a small number of relatively simple technology applications involved in deep learning. In general, these existing technologies have the following shortcomings:

(1) The input of the above method needs to be of fixed length to extract the characteristic parameters of fixed length. However, in the actual application scenario, the lung sound signal of variable length is obtained, and real-time signal acquisition and diagnosis are very important;

(2) There are also many types of rales, and different types of rales correspond to different diseases. Therefore, it is important to be able to identify different types of rales, but the above method does not give a multi-classification scheme for different types of rales. ；

(3) The conditions of the lung lesions of each patient are different, so that even the same kind of rales may present different lung sounds at different times. The existing technology is still very poor in the robustness of the rale detection and classification results.

With the rapid development of Internet of Things (IoT) technology and artificial intelligence (AI) technology in recent years, it is possible to classify lung rales in real time based on artificial intelligence methods. Therefore, it becomes an urgent need to realize a real-time classification method of lung rales of an electronic stethoscope.

Summary of the invention

The purpose of the present invention is to provide a method, system, device and readable storage medium for real-time artificial intelligence classification of lung rales of an electronic stethoscope, so as to solve the problems raised in the background art.

In order to achieve the above objectives, the present invention provides the following technical solutions:

An artificial intelligence real-time classification method for lung rales of an electronic stethoscope, including:

Step 1. From the start of lung sound acquisition by the electronic stethoscope, the data in the acquisition channel is read in real time to a certain buffer space. When the data is accumulated to 2 seconds, the automatic classification program of lung rales is started;

Step 2. Downsample the 2 second data block to f _s = 8kHz, pass through a band-pass filter, and normalize; if the data block is the i-th data block, count the preprocessed data block The data block is a vector x _i ;

Step 3. Calculate the logarithmic Mel filter bank transformation of the _{data vector x i} _{, expressed as a matrix F i} ;

Step 4. Use the logarithmic Mel filter bank transformation result matrix F _i to calculate the data matrix Δ _i,0 , Δ _i,1 and Δ _i,2 of the three channels;

Step 5. Normalize the data matrices Δ ₀ , Δ ₁ and Δ _{2 of} the three channels, and input a pre-built and trained convolutional neural network. The output of the convolutional neural network is four probability values: The probability that only wet rales exist in the data block p _i,c , the probability that only wheeze sounds in _{the data block p i,w} , the probability that both wet rales and wheeze exist in the data block p _i,cw , The data block does not have the probability p _{i,Null of} wet rales and stridor, save these four probability values p _i =[pi _,c ,pi _,w ,pi _,cw ,pi _,Null ] ^T.

Step 6. When the length of the data stored in the cache space reaches 3.9 seconds, remove the previous 1.9 seconds of data, use the remaining 2 seconds of data as the i+1 data block, and go back to step 2; when the data stored in the cache space When the lung sound collection is finished before the duration of 3.9 seconds, go to step 7;

Step 7. If the probability value of any data block is not saved in the end, the output will be "Cannot determine whether there is a rale"; if the probability values p ₁ ,p ₂ ,...,p _{N on N data blocks are finally saved} Using these probability values, output "only rales in lung sounds", "only wheezing sounds in lung sounds", "rales and stridor sounds in lung sounds at the same time" and "no rales in lung sounds" One of the four states of "tone", and the probability value of that state is given.

Preferably, the filter used in step 2 is a Butterworth band pass filter, and the pass band is 100 Hz to 1000 Hz.

Preferably, calculating the logarithmic Mel filter bank transformation matrix F _{i of the} _{data vector x i} in step 3 includes:

First, calculate x _i STFT spectrum: x _i will be divided into M = 31 sections each comprising N _FFT = 1024 sampling points, a 50% overlap between segments; paragraph order of data is represented by x m _i,m (n),n=0,1,...,N _FFT -1, then the fast Fourier transform of this segment is calculated as

Where h(n) is the Hamming window;

Then, |Y _i,m (k)| ² is filtered by a Mel filter bank; the Mel filter bank contains Q=29 Mel frequency domain ranges f _Mel (f) = 2959×log ₁₀ (1+ f/700),f～[0,f _s /2], uniformly spaced and 50% overlapping triangular filter Ψ _q ,q=1,2,...,Q; after filtering by the Mel filter bank The result is

Finally, the logarithmic Mel filter bank transformation matrix F _{i of} _{x i} is calculated, and the elements in the qth row and m column are given by the following formula: F _i [q,m]=log[y _i,m (q)] .

Preferably, the data matrix of the three channels calculated in step 4 includes:

First, the 29×29-dimensional data matrix on the first channel Δ ₀ =F[:,1:M-2];

Then, the 29×29-dimensional data matrix on the second channel Δ ₁ =F[:,2:M-1]-F[:,1:M-2];

Finally, the 29×29-dimensional data matrix Δ ₂ on the third channel = (F[:,3:M]-F[:,2:M-1])-Δ ₁ .

Preferably, the convolutional neural network in step 5 is trained from a large sample labeled data set. The specific structure of the network is shown in Figure 3. The convolutional neural network has 4 convolutional layers, and its convolution kernel The sizes are 5×5, 3×3, 3×3, and 3×3; the convolutional layer uses ReLU as the activation function; the pooling layer uses maximum pooling; the output layer outputs 4 probabilities p _{i, c} , p through softmax _i,w , p _i,cw and p _i,Null ; in the process of training the convolutional neural network, a truncated normal distribution with a standard deviation of 0.1 is used for the initial weight of the parameters, and Adam optimization, dropout learning and L _{2 are used at the same time} Regularization.

Preferably, the probability values corresponding to the four states that may finally be output in step 7 are:

Probability of "There are only rales in lung sounds"

Probability of "there is only wheezing in lung sounds"

Probability of "no rales in lung sounds"

The probability that "wet rales and wheezing sounds exist in lung sounds at the same time" is p _cw = 1-p _c- p _w- p _Null .

The final output is the state with the highest probability among the four states and its corresponding probability.

In order to solve the above technical problems, the present invention also provides an artificial intelligence real-time classification system for lung rales of an electronic stethoscope, including:

The electronic stethoscope collects lung sounds, allocates a buffer space for the collected data, and continuously enters the buffer. When the data accumulates to 2 seconds, the lung rales automatic classification program is started;

Band pass filter, filter the collected data and normalize it;

Logarithmic mel filter bank, transform the result matrix of the data vector to calculate the data matrix of the three channels;

Convolutional neural network, used for three-channel data matrix input, output and save four probability values;

Among them: electronic stethoscope, band pass filter, logarithmic mel filter bank and convolutional neural network are connected in sequence.

Preferably, the band-pass filter adopts a Butterworth band-pass filter with a pass band of 100 Hz to 1000 Hz. The convolutional neural network has 4 convolutional layers, and the size of the convolution kernel is 5×5, 3×3, and 3 respectively. ×3 and 3×3; the convolutional layer uses ReLU as the activation function; the pooling layer uses maximum pooling; the output layer is output through softmax.

To solve the above technical problems, the present invention also provides a real-time artificial intelligence classification device for pulmonary rales of an electronic stethoscope, which is characterized in that it includes:

Memory, used to store computer programs;

The processor is used to implement the steps of any one of the above-mentioned methods for artificial intelligence real-time classification of lung rales of the electronic stethoscope when the computer program is executed.

To solve the above technical problems, the present invention also provides a computer-readable storage medium, characterized in that a computer program is stored on the computer-readable storage medium. The steps of an artificial intelligence real-time classification method for lung rales of an electronic stethoscope are described.

Compared with the prior art, the beneficial effects of the present invention are:

(1) The present invention selects data blocks slidingly over time and inputs them into a specific convolutional neural network for classification, and finally combines the classification results of all data blocks to obtain the final total rale classification result, without the need to preset the length of the input data. Realize the real-time automatic classification of rales, and the use of multi-time period joint rale classification can improve the robustness of the classification results;

(2) The present invention extracts the three-channel logarithmic mel filter bank transform feature as the input of the convolutional neural network;

(3) The present invention clearly provides a specific and effective convolutional neural network structure for lung rales classification, in which the convolutional layer is used to discover the deeper features of the input data, and the pool is added after the convolutional layer Layer to improve the fault tolerance of the network;

(4) In the process of training the convolutional neural network, the present invention adds a truncated normal distribution with a standard deviation of 0.1 for parameter weight initialization. At the same time, it uses Adam optimization, Dropout learning, and L2 regularization to prevent overfitting and improve the cost. Robustness of the method;

(5) The present invention can realize multi-classification of the four cases of wet rales, wheezing, and both of them and neither of them.

Description of the drawings

Figure 1 is a schematic diagram of breath sounds with different additional rales in the prior art;

Fig. 2 is a flowchart of a method for real-time artificial intelligence classification of lung rales of an electronic stethoscope according to the present invention;

Fig. 3 is a structure diagram of a convolutional neural network used for four classifications of a single data block according to the present invention;

4 is a characteristic diagram of the preprocessing and extraction of the present invention: (a) is an example diagram of the signal waveform of the original acquisition; (b) is an example diagram of the signal waveform after preprocessing a certain 2-second data block;

5 is a schematic diagram of the structure of the lung rale artificial intelligence real-time classification system of the electronic stethoscope of the present invention;

Fig. 6 is a schematic diagram of the structure of the lung rale artificial intelligence real-time classification device of the electronic stethoscope of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work shall fall within the protection scope of the present invention.

Please refer to Figures 1 to 6, the present invention provides a technical solution:

The lung sound signal is collected in real time through the electronic stethoscope, and a buffer space is allocated for the collected data and continues to enter the buffer. When the data accumulates to 2 seconds, start the lung rales automatic classification program (the data waveform example is shown in Figure 4(a)).

The data block with a duration of 2 seconds is down-sampled to f _s = 8 kHz, filtered by a Butterworth band-pass filter with a pass band of 100 Hz to 1000 Hz, and normalized. Figure 4(b) shows a 2-second data block after preprocessing. If the data block is the i-th data block, the preprocessed data block is counted as the vector x _i , and the data vector x is calculated mel logarithmic transform filter bank _i is denoted as _i matrix F., the specific process is: (1) first, calculate the short time Fourier transform of the spectrum of x _i: x _i will be divided into M = 31 sections, each Contains N _FFT ＝1024 sampling points, the overlap between segments is 50%; let the m- _{th segment data be expressed as x i,m} (n),n=0,1,...,N _FFT -1, then The fast Fourier transform is calculated as

Where h(n) is the Hamming window; (2) Then, |Y _i,m (k)| ² is filtered by a mel filter bank, which contains Q=29 mel frequency domain ranges f _Mel (f)=2959×log ₁₀ (1+f/700), a triangular filter with uniform spacing and 50% overlap on _{f～[0,f s} _{/2] Ψ q} , q=1,2,. ..,Q, the result of Mel filter bank filtering is

(3) Finally, the logarithmic Mel filter bank transformation matrix F _{i of} _{x i} is calculated, and the elements in the qth row and m column are given by the following formula: F _i [q,m]=log[y _i,m ( q)]. After obtaining the logarithmic mel filter bank transformation result matrix F _i , pass

Calculate the data matrix Δ _i,0 , Δ _i,1 and Δ _i,2 of the three channels. Normalize the data matrices Δ ₀ , Δ ₁ and Δ _{2 of} the three channels, and input a pre-built and trained convolutional neural network (shown in Figure 3). The output of the convolutional neural network is four. Probability values: the probability p _i,c _{of only wet rales in the data block, the probability p i,w} of only wheezing sounds in the data block, the probability that both wet rales and wheezing sounds exist in the data block p _i,cw _{, the probability p i,Null} that the data block does not have wet rales and wheezing sounds, save these four probability values p _i =[pi _,c ,pi _,w ,pi _,cw , p _i,Null ] ^T.

When the length of the data stored in the cache space reaches 3.9 seconds, the previous 1.9 seconds of data are removed, the remaining 2 seconds of data are regarded as the i+1th data block, and the above process is repeated. When the data saved in the buffer space does not reach 3.9 seconds before the lung sound collection ends, a judgment is made: if the probability value of any data block is not saved in the end, the output is "Cannot judge whether there is a rale"; if it is finally saved _{Probability values p 1} ,p ₂ ,...,p _N on N data blocks, calculate the probability of "only wet rales in lung sounds"

Probability of "there is only wheezing in lung sounds"

Probability of "Role and stridor in lung sounds at the same time" p _cw = 1-p _c -p _w -p _Null and probability of "No rales in lung sounds"

Comparing the magnitudes of these four probabilities, the state with the highest probability is the recognized state. Outputs "only wet rales in lung sounds", "only stridor sounds in lung sounds", and "wet rales in lung sounds at the same time." One of the four states of "tone and stridor" and "no rales in lung sounds", and the probability value of this state is given.

Using the 920 pieces of lung sound data provided by the International Conference on Biomedicine and Health Information (covering the four lung sound situations involved in the present invention, each piece of data has a variable length and a duration of 10 seconds to 90 seconds) and the applicant team 508 pieces of lung sound data collected by the pediatric department of several domestic hospitals (also covering the four lung sound situations involved in the present invention, each piece of data is more than 30 seconds in length), a total of 1,428 pieces of data are used as a lung sound database for neural network training And verification of classification effect. Take the 1071 pieces of data as the training set, divide it into a total of 14,524 data blocks according to the data block sliding selection method of the present invention, and extract their respective three-channel logarithmic Mel filter bank transform features according to the aforementioned method and mark them , To train the convolutional neural network. The specific structure of the network is shown in Figure 3; the convolutional neural network has 4 convolutional layers, and the convolution kernel sizes are 5×5, 3×3, 3×3, and 3×3; the convolutional layer uses ReLU is used as the activation function; the pooling layer uses maximum pooling; the output layer outputs 4 probabilities p _i,c , p _i,w , p _i,cw and p _i,Null through softmax; in the process of training the convolutional neural network , A truncated normal distribution with a standard deviation of 0.1 is used for the initial weights of the parameters, and Adam optimization, Dropout learning, and L ₂ regularization are also used. Finally, using the remaining 357 pieces of lung sound data as the test set, the rale classification accuracy of the lung sound data section of the final test set is 95.80%.

The present invention proposes a real-time artificial intelligence classification method for pulmonary rales of an electronic stethoscope. The main technical problems to be solved include:

(1) How to provide a unified real-time classification method of rales under the condition that the total length of the actual lung sound collection is uncertain; (2) Since different rales are related to different diseases, how to realize the multi-classification of rales ; (3) How to improve the robustness of rale detection and classification results.

The present invention, through the above method, (1) can provide a unified real-time classification method of rales under the condition that the total length of actual lung sound collection is uncertain; (2) the present invention can realize wet rales, stridor, Both include and neither include the multi-classification of these four cases; (3) The present invention can effectively improve the robustness of rale detection and classification results.

Specifically:

Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, and substitutions can be made to these embodiments without departing from the principle and spirit of the present invention. And variations, the scope of the present invention is defined by the appended claims and their equivalents.

Claims

An artificial intelligence real-time classification method for lung rales of an electronic stethoscope, which is characterized in that it comprises:

Step 1. From the start of lung sound acquisition by the electronic stethoscope, the data in the acquisition channel is read in real time to a certain buffer space. When the data is accumulated to 2 seconds, the automatic classification program of lung rales is started;

Step 2. Downsample the 2 second data block to f s = 8kHz, pass through a band-pass filter, and normalize; if the data block is the i-th data block, count the preprocessed data block The data block is a vector x i ;

Step 3. Calculate the logarithmic Mel filter bank transformation of the data vector x i , expressed as a matrix F i ;

Step 4. Use the logarithmic Mel filter bank transformation result matrix F i to calculate the data matrix Δ i,0 , Δ i,1 and Δ i,2 of the three channels;

Step 5. Normalize the data matrices Δ 0 , Δ 1 and Δ 2 of the three channels, and input a pre-built and trained convolutional neural network. The output of the convolutional neural network is four probability values: The probability p i,c that there is only a wet rale in the data block, the probability that there is only a wheezing sound in the data block p i,w , the probability that the data block has both a wet rale and a wheezing sound p i,cw , The data block does not have the probability p i,Null of wet rales and stridor, save these four probability values p i =[pi ,c ,pi ,w ,pi ,cw ,pi ,Null ] T.

Step 6. When the length of the data stored in the cache space reaches 3.9 seconds, remove the previous 1.9 seconds of data, use the remaining 2 seconds of data as the i+1 data block, and go back to step 2; when the data stored in the cache space When the lung sound collection is finished before the duration of 3.9 seconds, go to step 7;

Step 7. If the probability value of any data block is not saved in the end, the output will be "Cannot determine whether there is a rale"; if the probability values p 1 ,p 2 ,...,p N on N data blocks are finally saved Using these probability values, output "only rales in lung sounds", "only wheezing sounds in lung sounds", "rales and stridor sounds in lung sounds at the same time" and "no rales in lung sounds" One of the four states of "tone", and the probability value of that state is given.
An artificial intelligence real-time classification method for lung rales of an electronic stethoscope according to claim 1, wherein the filter used in step 2 is a Butterworth band pass filter with a pass band of 100 Hz to 1000 Hz.
An artificial intelligence real-time classification method for lung rales of an electronic stethoscope according to claim 1, characterized in that,

The calculation of the logarithmic Mel filter bank transformation matrix F i of the data vector x i in step 3 includes:

First, calculate x i STFT spectrum: x i will be divided into M = 31 sections each comprising N FFT = 1024 sampling points, a 50% overlap between segments; paragraph order of data is represented by x m i,m (n),n=0,1,...,N FFT -1, then the fast Fourier transform of this segment is calculated as
k=0,1,...,N FFT /2-1, where h(n) is the Hamming window;

Then, |Y i,m (k)| 2 is filtered by a Mel filter bank; the Mel filter bank contains Q=29 Mel frequency domain ranges f Mel (f) = 2959×log 10 (1+ f/700),f～[0,f s /2], uniformly spaced and 50% overlapping triangular filter Ψ q ,q=1,2,...,Q; after filtering by the Mel filter bank The result is
q=1,2,...,Q;

Finally, the logarithmic Mel filter bank transformation matrix F i of x i is calculated, and the elements in the qth row and m column are given by the following formula: F i [q,m]=log[y i,m (q)] .
An artificial intelligence real-time classification method for lung rales of an electronic stethoscope according to claim 1, characterized in that,

The data matrix of the three channels calculated in step 4 includes:

First, the 29×29-dimensional data matrix on the first channel Δ 0 =F[:,1:M-2];

Then, the 29×29-dimensional data matrix on the second channel Δ 1 =F[:,2:M-1]-F[:,1:M-2];

Finally, 29 × 29-dimensional data on a third channel matrix Δ 2 = (F [:, 3: M] - F [:, 2: M-1]) - Δ 1.
An artificial intelligence real-time classification method for lung rales of an electronic stethoscope according to claim 1, characterized in that,

The convolutional neural network in step 5 is trained from a large-sample labeled data set. The specific structure of the network is shown in Figure 3. The convolutional neural network has 4 convolutional layers, and the convolution kernel sizes are respectively 5×5, 3×3, 3×3 and 3×3; the convolutional layer uses ReLU as the activation function; the pooling layer uses maximum pooling; the output layer outputs 4 probabilities p i,c , p i,w through softmax , P i, cw and p i, Null ; in the process of training the convolutional neural network, a truncated normal distribution with a standard deviation of 0.1 is used for the initial weight of the parameters, and Adam optimization, dropout learning and L 2 regularization are used at the same time.
An artificial intelligence real-time classification method for lung rales of an electronic stethoscope according to claim 1, characterized in that,

The probability values corresponding to the four possible output states in step 7 are:

Probability of "There are only rales in lung sounds"

Probability of "there is only wheezing in lung sounds"

Probability of "no rales in lung sounds"

The probability that "wet rales and wheezing sounds exist in lung sounds at the same time" is p cw = 1-p c- p w- p Null .
An artificial intelligence real-time classification system for pulmonary rales of an electronic stethoscope, which is characterized in that it comprises:

The electronic stethoscope collects lung sounds, allocates a buffer space for the collected data and continuously enters the buffer. When the data accumulates to 2 seconds, the lung rales automatic classification program is started;

Band pass filter, filter the collected data and normalize it;

Logarithmic mel filter bank, transform the result matrix of the data vector to calculate the data matrix of the three channels;

Convolutional neural network, used for three-channel data matrix input, output and save four probability values;

Among them: electronic stethoscope, band pass filter, logarithmic mel filter bank and convolutional neural network are connected in sequence.
The lung rale artificial intelligence real-time classification system of an electronic stethoscope according to claim 1, wherein the band-pass filter adopts a Butterworth band-pass filter with a pass band of 100 Hz to 1000 Hz, and a convolutional neural network There are 4 convolutional layers in total, and the convolution kernel sizes are 5×5, 3×3, 3×3, and 3×3; the convolutional layer uses ReLU as the activation function; the pooling layer uses maximum pooling; the output layer passes softmax output.
An artificial intelligence real-time classification device for pulmonary rales of an electronic stethoscope, which is characterized in that it comprises:

Memory, used to store computer programs;

The processor is configured to implement the steps of the method for real-time artificial intelligence classification of lung rales of the electronic stethoscope according to any one of claims 1 to 6 when the computer program is executed.
A computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the electronic stethoscope according to any one of claims 1 to 6 is The steps of an artificial intelligence real-time classification method of lung rales.