Method and device for recognizing bowel sounds

Abstract

The present invention relates to the field of medical technology, and in particular to a method and device for recognizing bowel sounds. The method comprises: collecting physiological sound data and bowel sound data at a preset sampling rate; wherein the physiological sound data includes cough sounds, respiratory sounds, heart sounds, and murmurs; and the bowel sound data includes bowel sounds and intestinal murmurs; extracting corresponding acoustic features from the physiological sound data and the bowel sound data respectively; constructing a physiological sound recognition model; inputting the acoustic features of the physiological sound data into the physiological sound recognition model, training, and obtaining a physiological sound prediction model; freezing all parameters in the physiological sound recognition model, and adding a new feature classification block to replace the feature classification block to generate a bowel sound recognition model; inputting the acoustic features of the bowel sound data into the bowel sound recognition model to generate a bowel sound prediction model, so as to solve the problem of low bowel sound recognition accuracy in the prior art.

Description

Technical Field

[0001] The present invention relates to the field of medical technology, and in particular to a method and device for recognizing bowel sounds. Background Art

[0002] Bowel sounds refer to the intermittent sound of gas passing through water (or gurgling) produced by the movement of gas and fluid within the intestinal cavity during intestinal peristalsis. Bowel sounds reflect the movement of the small intestine and are an important indicator for detecting intestinal diseases. Auscultation of bowel sounds is currently a commonly used method for diagnosing intestinal diseases. With the development of deep learning technology, researchers are gradually using deep learning methods to diagnose bowel sounds. Current bowel sound recognition methods use deep neural networks for model training to improve model performance. However, deep neural networks have a large number of parameters and require a large amount of data support to train effective models.

[0003] Because collecting bowel sounds is difficult and data is relatively limited, or resources are limited, this poses a challenge to model training. Furthermore, bowel sound signals themselves are relatively weak and susceptible to noise, which can also affect model performance and lead to low bowel sound recognition accuracy. Therefore, when designing a bowel sound recognition model, it is important to comprehensively consider factors such as data volume, data quality, and model complexity. Summary of the Invention

[0004] In view of this, an object of the present invention is to provide a method and device for recognizing bowel sounds, so as to solve the problem of low bowel sound recognition accuracy in the prior art.

[0005] According to a first aspect of an embodiment of the present invention, a method for recognizing bowel sounds includes:

[0006] Collecting physiological sound data and bowel sound data at a preset sampling rate; wherein the physiological sound data includes cough sounds, breath sounds, heart sounds and murmurs; and the bowel sound data includes bowel sounds and intestinal murmurs;

[0007] extracting corresponding acoustic features from the physiological sound data and the bowel sound data respectively;

[0008] Constructing a physiological sound recognition model; wherein the physiological sound recognition model includes a feature extraction block and a feature classification block; the feature extraction block uses a left and right branch structure to extract different feature information of the acoustic features, and the different feature information is used to train the physiological sound recognition model;

[0009] Inputting the acoustic features of the physiological sound data into the physiological sound recognition model to train and obtain a physiological sound prediction model;

[0010] Freezing all parameters in the physiological sound prediction model and adding a new feature classification block to replace the feature classification block to generate a bowel sound recognition model;

[0011] The acoustic features of the bowel sound data are input into the bowel sound recognition model to train and obtain a bowel sound prediction model.

[0012] Furthermore, before extracting corresponding acoustic features from the physiological sound data and the bowel sound data, the method further includes performing data preprocessing on the physiological sound data and the bowel sound data, including:

[0013] performing signal filtering on the physiological sound data and the bowel sound data to generate a physiological sound signal and a bowel sound signal;

[0014] The physiological sound signal and the bowel sound signal are subjected to frame processing and windowing processing to generate frequency domains of each frame of the physiological sound signal and each frame of the bowel sound signal.

[0015] Furthermore, extracting corresponding acoustic features from the physiological sound data and the bowel sound data respectively includes:

[0016] The corresponding acoustic features are extracted according to the frequency domain of each frame of physiological sound signal and each frame of bowel sound signal.

[0017] Furthermore, before constructing the physiological sound recognition model, the method further includes:

[0018] Performing data mean normalization on the acoustic features of the physiological sound data.

[0019] Furthermore, the construction of the physiological sound recognition model includes:

[0020] The different feature information is combined and processed and then input into a feature classification block to construct and generate the physiological sound recognition model;

[0021] The feature classification block includes a fully connected layer, and the fully connected layer is activated using a preset function.

[0022] Furthermore, the training process of the physiological sound recognition model includes:

[0023] Acquire a preset first number of data samples as first training samples; wherein each data sample includes: an acoustic feature of target physiological sound data and a category of the target physiological sound data;

[0024] The physiological sound recognition model is trained based on the data samples to obtain the physiological sound prediction model.

[0025] Further, including:

[0026] Acquire a preset second number of data samples as first verification data; wherein each data sample includes: an acoustic feature of the target verification physiological sound data and a category of the target verification physiological sound data;

[0027] Inputting the verification data into the physiological sound prediction model to obtain a prediction result;

[0028] Calculating a fit between the prediction result and a probability value of the category to which the corresponding physiological sound data actually belongs;

[0029] If the degree of fit is lower than a preset value, the physiological sound prediction model is retrained.

[0030] Furthermore, the training process of the bowel sound prediction model includes:

[0031] Acquire a preset third number of data samples as second training samples; wherein each data sample includes: an acoustic feature of the target bowel sound data and a category of the target bowel sound data;

[0032] The bowel sound recognition model is trained based on the third number of data samples to obtain the bowel sound recognition model.

[0033] Further, including:

[0034] Acquire a preset fourth number of data samples as second verification data; wherein each data sample includes: an acoustic feature of the target verification bowel sound data and a category of the target verification bowel sound data;

[0035] Inputting the verification data into the bowel sound prediction model to obtain a bowel sound prediction result;

[0036] Calculating the degree of fit between the bowel sound prediction result and the probability value of the actual category of the corresponding bowel sound data;

[0037] If the degree of fit is lower than a preset value, the bowel sound prediction model is retrained.

[0038] According to a second aspect of an embodiment of the present invention, a bowel sound recognition device includes:

[0039] An acquisition module, configured to acquire physiological sound data and bowel sound data at a preset sampling rate; wherein the physiological sound data includes cough sounds, breath sounds, heart sounds, and murmurs; and the bowel sound data includes bowel sounds and intestinal murmurs;

[0040] a processing module, configured to extract corresponding acoustic features from the physiological sound data and the bowel sound data respectively;

[0041] A construction module for constructing a physiological sound recognition model; wherein the physiological sound recognition model includes a feature extraction block and a feature classification block; the feature extraction block uses a left and right branch structure to extract different feature information of the acoustic features, and the different feature information is used to train the physiological sound recognition model;

[0042] a training module, configured to input the acoustic features of the physiological sound data into the physiological sound recognition model to train and obtain a physiological sound prediction model;

[0043] An adjustment module, configured to freeze all parameters in the physiological sound prediction model, and add a new feature classification block to replace the feature classification block, thereby constructing and obtaining a bowel sound recognition model;

[0044] A generation module is used to input the acoustic features of the bowel sound data into the bowel sound recognition model to train and obtain a bowel sound prediction model.

[0045] The technical solutions provided by the embodiments of the present invention may have the following beneficial effects:

[0046] The present invention uses physiological sound data to train a physiological sound recognition model to obtain a physiological sound prediction model, wherein the physiological sound prediction model includes a feature extraction block and a feature classification block. The acoustic features of the learned physiological sound data are retained by freezing all parameters in the physiological sound prediction model, and the feature classification block of the physiological sound recognition model is adjusted, and the feature classification block is replaced with a new feature classification block, which is more suitable for extracting the acoustic features of bowel sound data and constructing a bowel sound recognition model. In this way, fine-tuning of the physiological sound recognition model to the bowel sound recognition model is achieved. Finally, the acoustic features of the bowel sound data are input into the constructed bowel sound recognition model and trained. The final bowel sound recognition model can improve the accuracy of bowel sound data recognition, thereby avoiding the problem of requiring a large amount of bowel sound data training to generate a bowel sound recognition model with higher accuracy, and realizing a low-resource bowel sound recognition method.

[0047] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0049] Figure 1 is a flow chart illustrating a method for recognizing bowel sounds according to an exemplary embodiment;

[0050] Figure 2The figure is a schematic block diagram of a bowel sound recognition device according to an exemplary embodiment. DETAILED DESCRIPTION

[0051] Exemplary embodiments will be described in detail herein, examples of which are illustrated in the accompanying drawings. In the following description, when referring to the drawings, like numbers in different figures represent like or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all possible embodiments consistent with the present invention. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present invention, as detailed in the appended claims.

[0052] refer to Figure 1 , Figure 1 FIG. 1 is a flow chart of a method for recognizing bowel sounds according to an exemplary embodiment. Figure 1 As shown, the method may specifically include the following steps:

[0053] Step S11: collecting physiological sound data and bowel sound data at a preset sampling rate; wherein the physiological sound data includes cough sounds, breath sounds, heart sounds and murmurs; and the bowel sound data includes bowel sounds and intestinal murmurs;

[0054] Exemplary physiological sound data include the Coswara dataset, which contains breath sound and cough sound samples, each lasting 30 seconds and 20 seconds, totaling approximately 75 hours; the ICBHI_2017 dataset, which contains breath sound and non-breath sound samples, each lasting 30 to 35 seconds, totaling approximately 9.5 hours; the COUGHVID dataset, which contains cough sound and non-cough sound samples, each lasting 4 to 20 seconds, totaling approximately 130 hours; and the MIMIC-III Waveform dataset, which contains heart sound and non-heart sound samples, each lasting several hours to several days, totaling over 3,000 hours. The Bowel Sound dataset contains bowel sound and murmur samples, totaling approximately 5 hours. These physiological sound datasets contain four categories: cough sounds, breath sounds, heart sounds, and murmurs, with non-heart sounds, non-cough sounds, and non-breath sounds combined into the murmur category. The bowel sound dataset contains two categories: bowel sounds and murmurs.

[0055] Step S12: extracting corresponding acoustic features from the physiological sound data and the bowel sound data respectively;

[0056] Preferably, before extracting corresponding acoustic features from the physiological sound data and the bowel sound data, the method further includes performing data preprocessing on the physiological sound data and the bowel sound data, including:

[0057] Signal filtering is performed on the physiological sound data and the bowel sound data to generate a physiological sound signal and a bowel sound signal.

[0058] The physiological sound signal and the bowel sound signal are subjected to frame processing and windowing processing to generate frequency domains of each frame of the physiological sound signal and each frame of the bowel sound signal.

[0059] The data preprocessing process is as follows:

[0060] Step 1. Signal filtering

[0061] Adjust the sampling rate of the physiological sound dataset to f s After the signal is 22050Hz, it is filtered by a third-order Butterworth bandpass filter.

[0062]

[0063] Selectively retain the frequency signal in the range of 50Hz-3000 Hz and filter out other frequency signals. c is the cutoff frequency, s is the Laplace transform factor, and replacing s with z-1 gives the differential equation of the third-order Butterworth digital filter in the discrete time domain.

[0064]

[0065] Where x(n) is the input signal, i.e., the physiological sound signal that needs to be filtered, and y(n) is the output signal, i.e., the physiological sound signal that retains the frequency range of 50 Hz-3000 Hz.

[0066] a0, a1, a2, a3, b0, b1, b2, and b3 are the coefficients of a third-order Butterworth digital filter.

[0067] Adjust the sampling rate of the bowel sound dataset to f s After the signal is reduced to 22050 Hz, a third-order Butterworth band-pass filter is used to filter the signal, selectively retaining the frequency signal in the range of 60 Hz-1200 Hz and filtering out other frequency signals.

[0068] Step 2. Framing and windowing

[0069] According to step 1, the physiological sound data and bowel sound data after signal filtering are obtained, and the physiological sound data and bowel sound data are segmented according to the time length T seconds. When a data segment with a time length less than T seconds appears, it is determined whether the data segment is greater than seconds. If the data segment is larger than seconds, directly fill the data segment with 0 to fill the time length to T seconds. If the data segment is less than or equal to After obtaining the segmented physiological sound data and bowel sound data, short-time Fourier transform is applied to each segment of physiological sound data and bowel sound data.

[0070] N s =f s ×T

[0071]

[0072] Get the frequency domain representation x of each segment of physiological sound data and bowel sound data respectively n (k). Wherein, d(n) is the time domain representation of each segment of physiological sound data or bowel sound data, k is the different frequency values ​​in the frequency of each segment of physiological sound data or bowel sound data, n is the index value of each segment of physiological sound data or bowel sound data, N s Indicates the number of sample points of a certain segment of T seconds of physiological sound data or bowel sound data.

[0073] After obtaining the frequency domain representation of each segment of physiological sound data and bowel sound data, the frequency domain representation of each segment of physiological sound data and bowel sound data is framed and windowed using a Hamming window.

[0074]

[0075] where X nf (k f ) represents each frame of physiological sound data or bowel sound data obtained after performing frame and window operations on each segment of physiological sound data or bowel sound data, that is, the frequency value in the nfth frame of a segment of physiological sound data or bowel sound data is k f Frequency domain representation, k f are the different frequency values ​​in the frequency domain of each frame of physiological sound data or bowel sound data. w(m) is the Hamming window function.

[0076] In step S12, corresponding acoustic features are extracted from the physiological sound data and the bowel sound data, respectively. The process of extracting acoustic features is as follows:

[0077] Step 3:

[0078] Step A: Based on the frequency domain representation of each frame of physiological sound data and bowel sound data obtained in step 2, the power spectrum density value of each frame of physiological sound data and bowel sound data is calculated to describe the energy distribution of physiological sound data and bowel sound data at different frequency points.

[0079]

[0080] Among them, |X nf (k f )| means taking the modulus of the value, Pnf (k f ) represents the frequency value k in the nfth frame of a certain physiological sound data or bowel sound data f The power spectral density value of .

[0081] Step B: Convert the power spectrum density value of each frame of physiological sound data and bowel sound data into decibel units, and arrange the power spectrum density value of each frame of physiological sound data in chronological order into a two-dimensional matrix, so as to obtain the logarithmic spectrum graph X of a certain segment of physiological sound data. phs , the dimension is (N p ,N c ), similarly, the power spectrum density values ​​of each frame of bowel sound data are arranged in time sequence into a two-dimensional matrix, that is, the logarithmic spectrum diagram X of a certain segment of bowel sound data is obtained. bs , the dimension is (N p ,N c ).

[0082]

[0083]

[0084] where N c Indicates the number of frames of a certain segment of T seconds of physiological sound data or bowel sound data. Indicates that the value is rounded. N s Indicates the number of sampling points of a certain period of T seconds of physiological sound data or bowel sound data. N f It is represented by the frame length, that is, the sampling point length of each frame of physiological sound data or bowel sound data. d is represented by the frame shift, that is, the number of sampling points of each shift of the Hamming window function. N p Indicates the number of different frequencies in a segment of physiological sound data or bowel sound data.

[0085] Step C: Normalize the logarithmic spectrograms of all physiological sound data and bowel sound data. For example, suppose there is a physiological sound data or bowel sound data containing N logarithmic spectrograms, calculate the mean μ and standard deviation σ of the logarithmic spectrograms of all physiological sound data or bowel sound data

[0086]

[0087]

[0088]

[0089] Among them, X is the logarithmic spectrum before mean normalization, X ′ It is the logarithmic spectrum after mean normalization.

[0090] It is worth noting that the logarithmic spectrum diagram is the acoustic feature mentioned in step S12.

[0091] Step S13: Construct a physiological sound recognition model; wherein, the physiological sound recognition model includes a feature extraction block and a feature classification block; the feature extraction block uses a left and right branch structure to extract different feature information of the acoustic features, and the different feature information is used to train the physiological sound recognition model.

[0092] It is worth noting that the constructed physiological sound recognition model adopts a two-branch structure. The left branch uses VGGNet, which is mainly composed of two blocks. The first is the convolutional layer block, which uses multiple repeated convolutional layer blocks and adds a pooling layer after each convolutional layer block to reduce the size of the feature map. The depth and pooling size of different convolutional layer blocks may be different. The second is the fully connected layer block, which is followed by several fully connected layers. The feature map output by the global average pooling layer in the last layer of the branch is converted into a one-dimensional, length L V The right branch uses ResNet, which is mainly composed of two blocks. One is the residual block, which is the core part of ResNet. Each residual block contains two convolutional layers and a residual connection. The residual connection adds the input data directly to the output data. The second is the residual network. ResNet consists of multiple residual blocks and uses the method of stacking multiple residual blocks to build a deep model. The feature map output by the last layer of the branch is converted into a one-dimensional, length L through the global average pooling layer. R The left branch and the right branch are collectively referred to as feature extraction blocks. The vectors output by these two branches are concatenated in the last dimension through the concatenate layer to form a one-dimensional vector with a length of L. V +L R The vector of length L is obtained by merging the two branches. V +L R The feature vector is input into the feature classification block,

[0093] The feature classification block includes a fully connected layer and uses a softmax activation function. Through the above construction, a physiological sound recognition model is obtained.

[0094] In some specific embodiments, the physiological sound recognition model adopts a multi-branch structure. One branch uses VGGNet, which can effectively retain more feature information in the physiological sound signal. Another branch uses ResNet, which can effectively extract feature information from the physiological sound signal, and the feature information can be multi-scale features.

[0095] Step S14: inputting the acoustic features of the physiological sound data into the physiological sound recognition model to train and obtain a physiological sound prediction model;

[0096] It can be understood that step S14 is the training of the physiological sound recognition model.

[0097] Preferably, the training process of the physiological sound recognition model includes:

[0098] Acquire a preset first number of data samples as first training samples; wherein each of the data samples includes: acoustic features of target physiological sound data and a category of the target physiological sound data;

[0099] The physiological sound recognition model is trained based on the data samples to obtain the physiological sound prediction model.

[0100] Acquire a preset second number of data samples as first verification data; wherein each data sample includes: an acoustic feature of the target verification physiological sound data and a category of the target verification physiological sound data;

[0101] Inputting the verification data into the physiological sound prediction model to obtain a prediction result;

[0102] Calculating a fit between the prediction result and a probability value of the category to which the corresponding physiological sound data actually belongs;

[0103] If the degree of fit is lower than a preset value, the physiological sound prediction model is retrained.

[0104] In some specific embodiments, assuming T d samples as the input of the physiological sound recognition model, wherein the sample here refers to the logarithmic spectrum of the physiological sound data after mean normalization, that is, the first number of data samples includes T d Logarithmic spectrogram of physiological sound data.

[0105] Specifically, suppose the i-th sample x () is the logarithmic spectrum of the target physiological sound data, and the input dimension of the target physiological sound data is (N p ,N c ), the true label corresponding to the i-th sample is y () ∈0,1,2,…,K, where K represents the number of categories. The true label refers to the specific category to which a certain segment of physiological sound data belongs. For example, the heart sounds in the physiological sound data are set to category 0, the breathing sounds are set to category 1, the cough sounds are set to category 2, and the murmurs are set to category 3, for a total of 4 categories.

[0106] The recognition result of the model output calculated by forward transmission is That is, it represents the probability distribution of the i-th sample for the four categories. The probability distribution here refers to the probability of the possible values ​​​​for the four categories, and the sum of the probabilities is 1.

[0107] Using multivariate cross entropy loss function

[0108]

[0109] Where, Represents the multivariate cross entropy loss function under the physiological sound recognition model parameter θ, which is used to measure the difference between the prediction result of the physiological sound recognition model and the true label. θ is the physiological sound recognition model parameter, N L Represents the number of log spectrograms of physiological sound data. It is represented as the value of the jth category in the true label of sample i. This value can be 0 or 1, indicating that the sample belongs to the jth category (1) or does not belong to the jth category (0). represents the probability value predicted by the physiological sound recognition model for sample i belonging to the jth category. This value represents the probability predicted by the physiological sound recognition model for this sample belonging to the jth category. It can be understood that the above method of using the multivariate cross-entropy loss function to measure the degree of difference between the physiological sound recognition model's prediction and the true label is the verification component of physiological sound recognition model training.

[0110] In each iteration, the parameters θ of the physiological sound recognition model are updated using stochastic gradient descent.

[0111]

[0112] Where α represents the learning rate, Represents the gradient of the loss function to the physiological sound recognition model parameter θ. For a physiological sound sample (x () ,y () ), the corresponding loss function gradient is

[0113]

[0114] in represents the error between the prediction result of the physiological sound recognition model and the true label, x () Represents the input physiological sound samples.

[0115] After each iteration, the loss values ​​of all samples are calculated again using the new parameters, and the specified number of iterations are repeated until the model converges to obtain a trained physiological sound recognition model.

[0116] It can be understood that through a large amount of data on four types of physiological sounds, namely cough sounds, heart sounds, respiratory sounds and murmurs, an accurate physiological sound recognition model is constructed based on the acoustic features by extracting the acoustic features of the physiological sounds, and the physiological sound recognition model is trained to obtain a physiological sound prediction model.

[0117] Step S15: Freeze all parameters in the physiological sound prediction model, and add a new feature classification block to replace the feature classification block to construct and obtain a bowel sound recognition model;

[0118] Specifically, the bowel sound recognition model consists of two parts: a feature extraction block and a feature classification block. The feature extraction block in the bowel sound recognition model uses the same model structure as the feature extraction block in the physiological sound recognition model, retaining the VGGNet and ResNet branches, as well as the concatenate layer following the two branches.

[0119] A new feature classification block is added, which contains two fully connected layers and uses the softmax activation function to map it to two categories. The model finally obtains the probability distribution for the two categories. The probability distribution here refers to the probabilities of the possible values ​​​​for the two categories, and the sum of the probabilities is 1.

[0120] The feature classification block of the physiological sound recognition model is replaced with the newly constructed feature classification block. After the above model construction, a bowel sound recognition model is obtained.

[0121] It can be understood that by utilizing the similarity of acoustic features between physiological sounds and bowel sounds, the obtained physiological sound prediction model is fine-tuned to obtain the required bowel sound recognition model. In this way, a bowel sound prediction model that can recognize bowel sounds with a high accuracy can be obtained without the need for a large amount of bowel sound data.

[0122] Step S16: Input the acoustic features of the bowel sound data into the constructed bowel sound recognition model, train and obtain a bowel sound prediction model.

[0123] Preferably, the training process of the bowel sound prediction model includes:

[0124] Acquire a preset third number of data samples as second training samples; wherein each of the data samples includes: an acoustic feature of the target bowel sound data and a category of the target bowel sound data;

[0125] The bowel sound recognition model is trained based on the third number of data samples to obtain the bowel sound prediction model.

[0126] Preferably, including:

[0127] Acquire a preset fourth number of data samples as second verification data; wherein each data sample includes: an acoustic feature of the target verification bowel sound data and a category of the target verification bowel sound data;

[0128] Inputting the verification data into the bowel sound prediction model to obtain a bowel sound prediction result;

[0129] Calculating the degree of fit between the bowel sound prediction result and the probability value of the actual category of the corresponding bowel sound data;

[0130] If the degree of fit is lower than a preset value, the bowel sound prediction model is retrained.

[0131] Specifically:

[0132] Step a: Assume that the third number of data samples is T t training samples, T t Training samples are used as input to the bowel sound recognition model, where the training samples refer to the logarithmic spectrogram of the target bowel sound data after mean normalization.

[0133] The i-th sample The input dimension is (N p ,N c ), the corresponding true label is The true label refers to the specific category to which a certain segment of bowel sound data belongs. For example, intestinal murmurs are set to category 0 and bowel sounds are set to category 1.

[0134] Step b: The prediction result of the model output is calculated through forward transmission. That is, it represents the probability distribution of the i-th sample for two categories. The probability distribution here refers to the probability of the two categories taking values, and the sum of the probabilities is 1. Using the binary cross entropy loss function

[0135]

[0136] Where, represents the bowel sound recognition model parameter θ b The binary cross entropy loss function under θ is used to measure the difference between the predicted results of the bowel sound recognition model and the true label, θ b is the parameter of the bowel sound recognition model, N represents the number of logarithmic spectrograms of bowel sound data, Indicates the true label of the i-th bowel sound sample, this value can be 0 or 1, Represents the probability value of the predicted output of the bowel sound recognition model for sample i.

[0137] Step c: Use the back propagation algorithm to calculate the loss function for the bowel sound recognition model parameters θ bGradient

[0138]

[0139] In each iteration, the stochastic gradient descent method is used to update the bowel sound recognition model parameters θ b

[0140]

[0141] Here, α represents the learning rate.

[0142] Repeat steps a to c until the model converges, and finally obtain a trained bowel sound recognition model.

[0143] It can be understood that the present invention uses physiological sound data to train a physiological sound recognition model to obtain a physiological sound prediction model, wherein the physiological sound prediction model includes a feature extraction block and a feature classification block. The acoustic features of the learned physiological sound data are retained by freezing all parameters in the physiological sound prediction model, and the feature classification block of the physiological sound recognition model is adjusted, and the feature classification block is replaced with a new feature classification block, which is more suitable for extracting the acoustic features of bowel sound data and constructing a bowel sound recognition model. In this way, fine-tuning of the physiological sound recognition model to the bowel sound recognition model is achieved. Finally, the acoustic features of the bowel sound data are input into the constructed bowel sound recognition model and trained. The final bowel sound recognition model can improve the accuracy of bowel sound data recognition, thereby avoiding the problem of requiring a large amount of bowel sound data training to generate a bowel sound recognition model with higher accuracy, and realizing a low-resource bowel sound recognition method.

[0144] Reference Figure 2 , Figure 2 is a schematic block diagram of a bowel sound recognition device, such as Figure 2 Shown, including:

[0145] Acquisition module 1, for collecting physiological sound data and bowel sound data at a preset sampling rate; wherein the physiological sound data includes cough sounds, breath sounds, heart sounds and murmurs; the bowel sound data includes bowel sounds and intestinal murmurs;

[0146] Processing module 2, for extracting corresponding acoustic features from the physiological sound data and the bowel sound data respectively;

[0147] Construction module 3 is used to construct a physiological sound recognition model; wherein the physiological sound recognition model includes a feature extraction block and a feature classification block; the feature extraction block uses a left and right branch structure to extract different feature information of the acoustic features, and the different feature information is used to train the physiological sound recognition model;

[0148] A training module 4 is configured to input the acoustic features of the physiological sound data into the physiological sound recognition model to train and obtain a physiological sound prediction model;

[0149] An adjustment module 5 is configured to freeze all parameters in the physiological sound prediction model, and add a new feature classification block to replace the feature classification block, thereby constructing and obtaining a bowel sound recognition model;

[0150] The generating module 6 is used to input the acoustic features of the bowel sound data into the bowel sound recognition model to train and obtain a bowel sound prediction model.

[0151] Specifically, a bowel sound recognition device can refer to the specific implementation of a bowel sound recognition method described in any of the above embodiments, which will not be repeated here.

[0152] It can be understood that the present invention uses physiological sound data to train a physiological sound recognition model to obtain a physiological sound prediction model, wherein the physiological sound prediction model includes a feature extraction block and a feature classification block. The acoustic features of the learned physiological sound data are retained by freezing all parameters in the physiological sound prediction model, and the feature classification block of the physiological sound recognition model is adjusted, and the feature classification block is replaced with a new feature classification block, which is more suitable for extracting the acoustic features of bowel sound data and constructing a bowel sound recognition model. In this way, fine-tuning of the physiological sound recognition model to the bowel sound recognition model is achieved. Finally, the acoustic features of the bowel sound data are input into the constructed bowel sound recognition model and trained. The final bowel sound recognition model can improve the accuracy of bowel sound data recognition, thereby avoiding the problem of requiring a large amount of bowel sound data training to generate a bowel sound recognition model with higher accuracy, and realizing a low-resource bowel sound recognition method.

[0153] It can be understood that the same or similar parts of the above embodiments can be referenced to each other, and the contents not described in detail in some embodiments can refer to the same or similar contents in other embodiments.

[0154] It should be noted that, in the description of the present invention, the terms "first", "second", etc. are used for descriptive purposes only and should not be understood as indicating or implying relative importance. In addition, in the description of the present invention, unless otherwise specified, the meaning of "plurality" is at least two.

[0155] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code comprising one or more executable instructions for implementing the steps of a specific logical function or process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may be performed out of the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present invention pertain.

[0156] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used: logic gate circuits and discrete logic circuits for implementing logical functions on digital signals, application-specific integrated circuits with suitable combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc.

[0157] Those skilled in the art will understand that all or part of the steps in the method of the above embodiment can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium. When the program is executed, it includes one or a combination of the steps of the method embodiment.

[0158] In addition, the functional units in the various embodiments of the present invention may be integrated into a single processing module, or each unit may exist physically separately, or two or more units may be integrated into a single module. The aforementioned integrated modules may be implemented in the form of hardware or in the form of software functional modules. If the integrated modules are implemented in the form of software functional modules and sold or used as independent products, they may also be stored in a computer-readable storage medium.

[0159] The storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disk, etc.

[0160] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that the specific features, results, materials, or characteristics described in conjunction with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, results, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0161] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A method for recognizing bowel sounds, characterized in that: include: Collecting physiological sound data and bowel sound data at a preset sampling rate; wherein the physiological sound data includes cough sounds, breath sounds, heart sounds and murmurs; and the bowel sound data includes bowel sounds and intestinal murmurs; extracting corresponding acoustic features from the physiological sound data and the bowel sound data respectively; Constructing a physiological sound recognition model; wherein the physiological sound recognition model includes a feature extraction block and a feature classification block; the feature extraction block uses a left and right branch structure to extract different feature information of the acoustic features, and the different feature information is used to train the physiological sound recognition model; Inputting the acoustic features of the physiological sound data into the physiological sound recognition model to train and obtain a physiological sound prediction model; Freezing all parameters in the physiological sound prediction model and adding a new feature classification block to replace the feature classification block to generate a bowel sound recognition model; The acoustic features of the bowel sound data are input into the bowel sound recognition model to train and obtain a bowel sound prediction model.

2. The method according to claim 1, characterized in that Before extracting corresponding acoustic features from the physiological sound data and the bowel sound data, the method further includes performing data preprocessing on the physiological sound data and the bowel sound data, including: performing signal filtering on the physiological sound data and the bowel sound data to generate a physiological sound signal and a bowel sound signal; The physiological sound signal and the bowel sound signal are subjected to frame processing and windowing processing to generate frequency domains of each frame of the physiological sound signal and each frame of the bowel sound signal.

3. The method according to claim 1, characterized in that The extracting corresponding acoustic features from the physiological sound data and the bowel sound data respectively includes: The corresponding acoustic features are extracted according to the frequency domain of each frame of physiological sound signal and each frame of bowel sound signal.

4. The method according to claim 1, wherein Before constructing the physiological sound recognition model, the method further includes: Performing data mean normalization on the acoustic features of the physiological sound data.

5. The method according to claim 1, wherein The step of constructing a physiological sound recognition model comprises: The different feature information is combined and processed and then input into a feature classification block to construct and generate the physiological sound recognition model; The feature classification block includes a fully connected layer, and the fully connected layer is activated using a preset function.

6. The method according to claim 4, characterized in that The training process of the physiological sound recognition model includes: Acquire a preset first number of data samples as first training samples; wherein each data sample includes: an acoustic feature of target physiological sound data and a category of the target physiological sound data; The physiological sound recognition model is trained based on the data samples to obtain the physiological sound prediction model.

7. The method according to claim 6, characterized in that include: Obtaining a preset second number of data samples as first verification data; Each data sample includes: acoustic features of target verification physiological sound data and categories of target verification physiological sound data; Inputting the verification data into the physiological sound prediction model to obtain a prediction result; Calculating a fit between the prediction result and a probability value of the category to which the corresponding physiological sound data actually belongs; If the degree of fit is lower than a preset value, the physiological sound prediction model is retrained.

8. The method according to claim 1, characterized in that The training process of the bowel sound prediction model includes: Acquire a preset third number of data samples as second training samples, wherein each data sample includes: an acoustic feature of the target bowel sound data and a category of the target bowel sound data; The bowel sound recognition model is trained based on the third number of data samples to obtain the bowel sound recognition model.

9. The method according to claim 8, characterized in that include: Obtaining a preset fourth number of data samples as second verification data; Each data sample includes: acoustic features of target verification bowel sound data and categories of target verification bowel sound data; Inputting the verification data into the bowel sound prediction model to obtain a bowel sound prediction result; Calculating the degree of fit between the bowel sound prediction result and the probability value of the actual category of the corresponding bowel sound data; If the degree of fit is lower than a preset value, the bowel sound prediction model is retrained.

10. A bowel sound recognition device, characterized in that: include: An acquisition module, configured to acquire physiological sound data and bowel sound data at a preset sampling rate; wherein the physiological sound data includes cough sounds, breath sounds, heart sounds, and murmurs; and the bowel sound data includes bowel sounds and intestinal murmurs; a processing module, configured to extract corresponding acoustic features from the physiological sound data and the bowel sound data respectively; A construction module for constructing a physiological sound recognition model; wherein the physiological sound recognition model includes a feature extraction block and a feature classification block; the feature extraction block uses a left and right branch structure to extract different feature information of the acoustic features, and the different feature information is used to train the physiological sound recognition model; a training module, configured to input the acoustic features of the physiological sound data into the physiological sound recognition model to train and obtain a physiological sound prediction model; An adjustment module, configured to freeze all parameters in the physiological sound prediction model, and add a new feature classification block to replace the feature classification block, thereby constructing and obtaining a bowel sound recognition model; A generation module is used to input the acoustic features of the bowel sound data into the bowel sound recognition model to train and obtain a bowel sound prediction model.

Images