Lung disease classification and detection method based on collaborative deep learning and lung auscultation sounds

The lung auscultation sound classification and detection method based on collaborative deep learning, utilizing residual neural networks and collaborative network models, solves the problem of poor diagnostic accuracy of lung auscultation sounds, achieves rapid and accurate diagnosis of lung diseases, and improves the accuracy and adaptability of diagnosis.

CN114869224BActive Publication Date: 2024-11-08ZHEJIANG UNIV +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210311257.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-03-28
Publication Date
2024-11-08
Estimated Expiration
2042-03-28

AI Technical Summary

Technical Problem

In existing technologies, the audio band for lung auscultation is within the range of human ear insensitivity, resulting in poor diagnostic accuracy and reliability. Furthermore, reliance on doctors' experience can easily lead to missed diagnoses and misdiagnoses, making it difficult to achieve rapid and accurate diagnosis of lung diseases.

Method used

A collaborative deep learning-based approach is adopted, utilizing a model consisting of two pre-trained residual neural networks and a collaborative network. The collaborative network model is trained end-to-end, and a lung disease classification and detection model is constructed through pre-emphasis, frame segmentation, fast Fourier transform, and feature extraction, thereby enhancing the network's feature representation capabilities.

Benefits of technology

It improves the ability to distinguish easily confused samples, enhances the accuracy and generalization of lung disease classification, is applicable to various environments and scenarios, reduces reliance on doctors' experience, and improves the timeliness and accuracy of diagnosis.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114869224B_ABST
    Figure CN114869224B_ABST
Patent Text Reader

Abstract

The present invention provides a lung disease classification and detection method based on collaborative deep learning and lung auscultation sounds, which adopts a collaborative learning method between two ResNet-50s and is trained through a paired learning mode. Each time the model accepts a data pair as input, a pair of audio feature values ​​are respectively transmitted to the corresponding ResNet-50; these ResNet-50s are initialized and trained by a fine-tuning pre-training model method, and a collaborative learning system is designed to enable the two ResNet-50s to learn from each other. The collaborative system is used to supervise the similarities and differences of the data pairs, that is, whether they belong to the same category, and to back-propagate the collaborative errors generated by the two ResNet-50s in real time and correct the network weights, thereby further enhancing the network's ability to learn feature representation and being able to more effectively make accurate judgments on easily confused samples.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of medical data classification detection, and specifically relates to a lung disease classification detection method based on collaborative deep learning and lung auscultation sounds. Background Art

[0002] The lung is a vital organ for maintaining the normal functioning of the body. It is connected to the outside world, with 10,000 liters of gas entering and exiting every day. It has a wide gas exchange area with the outside world, so the lung's defense function is very heavy. The lung not only carries out daily gas exchange activities, but also plays a great role in the body's metabolism and immune function. However, since the advent of the industrial society, with the continuous development of industrialization and economy, its byproducts have also caused certain damage to the environment. The number of patients with respiratory diseases has increased day by day, which is closely related to the increasingly severe air pollution, the increasing number of smokers and the widespread aging of the population. Nowadays, respiratory diseases have become a common and frequently occurring disease, and most of the lesions are in the upper trachea, bronchi, lungs and chest cavity; those with mild lesions often cough, have chest pain, and even suffer from severe damage to respiratory function; those with severe lesions die of ventilatory disorders, hypoxia, and even circulatory failure. Among the top ten causes of death in the country announced in 2006, respiratory diseases (excluding lung cancer) ranked fourth in urban causes of death, accounting for 13.1%, while in rural causes of death, it ranked third (16.4%).

[0003] Lung auscultation sounds are one of the common methods for diagnosing lung diseases. Compared with medical images that require professional equipment and operators to obtain, lung auscultation sounds have more flexible diagnostic scenarios and lower usage thresholds. Doctors can judge the type of disease by different respiratory sound characteristics such as alveolar breath sounds, bronchial sounds, and chest friction sounds. It has a good effect on the detection of early lung lesions. However, since the frequency band of lung auscultation sounds is between 100Hz and 2000Hz, it is not completely within the sensitive frequency range of the human ear. Therefore, it is easy for the human ear to lose important low-frequency information, affecting the accuracy and reliability of the diagnosis results. Moreover, the auscultation effect is closely related to the doctor's level and professional experience, and is also affected by the external environment and hearing conditions. For various reasons, relying on lung auscultation sounds to diagnose lung diseases may have the risk of missed diagnosis and misdiagnosis.

[0004] my country has a large population base and a very high incidence of lung diseases. Rapid and accurate diagnosis and treatment of lung diseases has always been a problem that my country urgently needs to face. The most common solution is to combine lung disease diagnosis with deep learning to realize an automatic disease classification model, share the diagnostic pressure of doctors, improve the timeliness and accuracy of lung disease diagnosis, and enable patients to detect lesions as early as possible and receive timely treatment. Using a digital lung stethoscope to obtain lung auscultation sounds as a diagnostic method has strong flexibility and can adapt to a variety of environmental scenarios. Combining deep learning technology with lung auscultation sound diagnosis can better promote the application of smart medical care and, to a certain extent, alleviate the current situation of difficulty in seeing a doctor in remote and backward areas. Summary of the invention

[0005] In view of the above, the present invention provides a lung disease classification and detection method based on collaborative deep learning and lung auscultation sounds. The model is composed of two pre-trained residual neural networks and one collaborative network, and the collaborative network model is trained end-to-end to further enhance the network's ability to learn feature representation, which can effectively make accurate judgments on easily confused samples.

[0006] A lung disease classification detection method based on collaborative deep learning and lung auscultation sounds comprises the following steps:

[0007] (1) obtaining a certain number of lung auscultation sounds with classification labels, and pre-emphasizing the data signals of the lung auscultation sounds to remove low-frequency noise in the signals;

[0008] (2) Framing and windowing the pre-emphasized data signal.

[0009] (3) Performing a fast Fourier transform on each frame of data signal to obtain the energy distribution of the signal in the frequency spectrum;

[0010] (4) Extract features from the spectrum of each frame of data signal, take the feature matrix of the entire data signal and its corresponding classification label as a group of samples, and then divide all samples into a training set and a test set;

[0011] (5) Construct a classification detection model consisting of two residual neural networks ResNet-50 and a collaborative network. The outputs of the two residual neural networks ResNet-50 are used as the input of the collaborative network.

[0012] (6) The training set samples are input into two residual neural networks ResNet-50 in pairs to train the model. Finally, the trained model is used to classify and detect the test set samples.

[0013] Furthermore, in the step (1), a high-pass filter is used to filter the data signal to remove low-frequency noise in the signal and increase the proportion of the high-frequency part, which can eliminate the inhibitory effect of the sound system on the high-frequency part and highlight the high-frequency resonance peak.

[0014] Furthermore, in step (2), a Hamming window is used for weighted framing to realize signal segmentation. When the window is shifted, only 1 / 2 or 1 / 3 of the window length is shifted, so that there is an overlap between two adjacent frames. The side lobe attenuation of the Hamming window is relatively large, which can highlight the data in the middle.

[0015] Furthermore, the specific implementation method of step (4) is as follows: first, the spectrum of the data signal is passed through a set of triangular filters to obtain the corresponding Mel frequency features, and then the Mel frequency features are used to calculate the logarithmic expression of the energy of each frame of data signal, and then the logarithmic energy of each frame of data signal is discrete cosine transformed to obtain the corresponding multi-order Mel feature cepstrum value to complete feature extraction; finally, the Mel feature cepstrum values ​​of the entire data signal are combined into the form of a feature matrix and the feature matrix is ​​channel-amplified to match the input requirements of the three channels of the model.

[0016] Furthermore, the loss function expression adopted by the residual neural network ResNet-50 is as follows:

[0017]

[0018] Where: L(θ) is the loss function of the residual neural network ResNet-50 under parameter θ, N is the number of samples input for each batch, K is the number of categories, It is the deep feature generated by the penultimate fully connected layer of the residual neural network ResNet-50 when the i-th sample in this batch is input, where the subscript j represents the category of the sample input to the residual neural network ResNet-50 detection output, l{y(i)=j} is the discriminant function, if y(i)=j holds, then l{y(i)=j}=1, otherwise l{y(i)=j}=0, y(i) is the true value category of the i-th sample.

[0019] Furthermore, the collaborative network is composed of an embedding layer, a fully connected learning layer and an output layer connected in sequence, and the deep features generated by the two residual neural networks ResNet-50 in the penultimate fully connected layer are represented as f A and f B , f A With f B Connect to get a combined feature f A,B The output is input into the collaborative network, and the output result is used to determine whether the sample pairs input into the two residual neural networks ResNet-50 belong to the same category. If there is a collaborative error, correction feedback is provided.

[0020] Furthermore, the loss function expression adopted by the collaborative network is as follows:

[0021]

[0022]

[0023] Where: L s (θ s ) is the collaborative network with parameter θ s The loss function under is the output of the collaborative network, y s is the label value of the collaborative network, y A and B is the true value category of the sample pair input to the two residual neural networks ResNet-50.

[0024] Furthermore, the present invention uses a small batch stochastic gradient descent algorithm to train the model according to the loss function of the residual neural network ResNet-50 and the collaborative network to optimize the model parameters.

[0025] Furthermore, when the trained model is used for testing, the test set samples are input into two residual neural networks ResNet-50 in pairs, and two corresponding prediction vectors P(A) and P(B) are obtained, which are the outputs of the last fully connected layer in the residual neural network ResNet-50. At this time, the additional collaborative network is discarded in the final classification prediction, and the final output prediction result is y:

[0026]

[0027] Where: M is the dimension of the prediction vector, P(A) i and P(B) i are the i-th element values ​​in the prediction vectors P(A) and P(B) respectively.

[0028] The present invention proposes a collaborative deep learning method for lung disease classification detection. The method simultaneously learns the representation of distinguishing features from multiple lung auscultation audio data, so that the ResNet-50 modules in the collaborative deep learning model can promote each other in the learning process (because if one ResNet-50 makes a correct decision, the error of another ResNet-50 may cause a collaborative error), thereby back-propagating the collaborative errors generated by the two ResNet-50s in real time and correcting the network weights, further enhancing the network's representation ability, and increasing the training weights of easily confused samples, so that they can be more accurately judged.

[0029] Due to the collaborative learning method adopted in the present invention, the feature learning capability of ResNet-50 is enhanced, and the problem of low accuracy of the algorithm based on deep learning is optimized; the added collaborative learning module enables ResNet-50 to learn from each other and optimize together, and the model has good discrimination ability for easily confused samples, which can greatly improve the ability to distinguish the categories of lung auscultation sounds and the classification accuracy of different lung diseases. The network structure used is suitable for three-channel image data. The present invention successfully trains on this network by amplifying one-dimensional sound data and simulating the consistent characteristics of the three-channel data values ​​of grayscale images. It can be proved that this network has good generalization and can be applied to different types of medical data. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 Schematic diagram of the extraction process of speech feature Mel-cepstral coefficients.

[0031] Figure 2 Schematic diagram of the audio signal waveform after framing, windowing and fast Fourier transform.

[0032] Figure 3 This is a schematic diagram of the principle of the Mel frequency filter bank.

[0033] Figure 4 The spectrograms are preprocessed and the Mel spectrum features are extracted. (a) is the original spectrogram, (b) is the spectrogram after high-pass filtering, and (c) is the spectrogram after Mel filtering.

[0034] Figure 5 This is a schematic diagram of the collaborative network model structure adopted by the present invention.

[0035] Figure 6 This is a schematic diagram of the structure of the collaborative network part. DETAILED DESCRIPTION

[0036] In order to describe the present invention more specifically, the technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0037] The present invention provides a lung disease classification detection method based on collaborative deep learning and lung breath sounds, which specifically includes the following steps:

[0038] (1) Data preprocessing.

[0039] In speech recognition, the most commonly used speech feature is the Mel-frequency cepstral coefficient. Due to the masking effect of the human ear, that is, the human ear is more sensitive to frequencies with high loudness and very insensitive to frequencies with low loudness, there is auditory sensitivity to signals of different frequencies. The distance that lower-frequency sound waves propagate on the basilar membrane of the inner cochlea is greater than that of higher-frequency sound waves, so the critical bandwidth from low frequency to high frequency is different. Specifically, a set of bandpass filters is applied to the frequency band from low frequency to high frequency, and this set of bandpass filters is arranged from dense to sparse according to the size of the critical bandwidth. After further processing, the signal energy output by each bandpass filter can be used as the input feature of speech.

[0040] Mel cepstral coefficients are cepstral parameters extracted in the Mel scale frequency domain. The Mel scale is the specific expression of linear frequency in human ear sensitivity. Its relationship with frequency can be approximately expressed by the following formula:

[0041]

[0042] Where: f is frequency, unit is Hz.

[0043] The specific steps for extracting Mel cepstral coefficients are as follows: Figure 1 As shown:

[0044] 1.1 Pre-emphasis

[0045] The speech signal passes through a high-pass filter to remove low-frequency noise, which is pre-emphasis:

[0046] H(z)=1-μz -1

[0047] Among them: the value of μ is 0.97; pre-emphasis can enhance the proportion of the high-frequency part, and the signal becomes flat after noise removal. At the same time, the inhibitory effect of the sound system on the high-frequency part is eliminated, and the high-frequency resonance peak can be highlighted.

[0048] 1.2 Framing

[0049] When processing audio signals, in order to extract more features from long-duration signals, the signals are usually framed, that is, the long sound signal is divided into multiple short sound signals of equal length. A short sound signal is called a frame, and each frame contains N sampling points, where the value of N is generally 512. Generally, there is an overlap between two adjacent frames to avoid too much change between the two adjacent frames. The overlapping area is generally 1 / 2 or 1 / 3 of N; for example, the sampling frequency of the lung auscultation sound signal is assumed to be aHz. For BMS data, the time length corresponding to each frame is

[0050] 1.3 Add Hamming Window

[0051] Framing is achieved by weighting a movable finite-length window on the audio data. Because adding a rectangular window will cause spectrum leakage, in order to avoid this phenomenon, the method of adding a Hamming window is adopted, because the side lobe attenuation of the Hamming window is large, which can highlight the middle data. When framing, the Hamming window is applied to each frame separately, and only 1 / 2 or 1 / 3 of the window is moved when the window is moved. Corresponding to the overlap of two adjacent frames, the data lost in the previous frames can be re-emphasized and used to increase the continuity of the left and right ends of each frame: S′(n)=S(n)×W(n), S(n) is the signal after the framing step, n=0,1…,N-1, W(n) is the Hamming window, and the form is as follows:

[0052]

[0053] Among them: a is the window size, n is the window number. The sampling frequency used in the experiment is 16000Hz, 1000ms, N is 512, and the overlapping area is N / 2, that is, 256. Therefore, after framing and windowing, a piece of audio can get 624 frames, each frame is 32ms long and contains 512 sampling points.

[0054] 1.4 Fast Fourier Transform

[0055] Since the energy distribution of the signal in the frequency domain can better reflect its characteristics, each frame needs to be fast Fourier transformed to obtain the energy distribution of the signal in the spectrum. In order to obtain the power spectrum, the spectrum of the speech signal needs to be squared modulo. The discrete Fourier transform of the speech signal is:

[0056]

[0057] Where: x(n) is the input signal, N represents the number of frames, that is, the number of points of Fourier transform; the audio signal after framing, windowing and fast Fourier transform is as follows Figure 2 shown.

[0058] 1.5 Triangular Bandpass Filter

[0059] The Mel frequency characteristics can be obtained by passing the processed signal spectrum through a set of triangular filter banks. The center frequency of the triangular filter bank is f(m), m = 1, 2, ..., M, M is the number of triangular filters, such as Figure 3 As shown, the interval between each f(m) widens as the value of m increases and shortens as m decreases.

[0060] The frequency response of a triangular filter is defined as:

[0061]

[0062] Where: After preprocessing and extraction of Mel spectrum features, the signal spectrum is as follows Figure 4 shown.

[0063] 1.6 Calculate the logarithmic expression of the energy output by each filter group respectively. The specific operation is:

[0064]

[0065] 1-7 is transformed by discrete cosine to obtain the coefficient C(n) required for input:

[0066]

[0067] The obtained logarithmic energy is subjected to discrete cosine transform to obtain the L-order Mel feature cepstrum value, where the L-order refers to the coefficient order of the Mel cepstrum coefficient, which is 15 here.

[0068] After the above series of processing, the original lung auscultation sound signal obtained the spectrum of the MFCC feature vector. In the process of obtaining the MFCC parameters, a Hamming window is used with a window length of 1024 and a spacing of 512, that is, 50% overlap between every two adjacent windows, and the number of Mel filters is 128. For sounds of 10 seconds in length, a 128*313 feature matrix can be obtained; in order to make it similar to the three channels of the model, the channel is amplified so that the training size of the input network is 218*313*3.

[0069] (2) Input data.

[0070] Different from the traditional ResNet-50, the collaborative deep learning proposed in the present invention simultaneously accepts n input data randomly selected from the training set. Each data and its class label are input into ResNet-50 together, and each pair of data has a corresponding collaborative label to determine whether the pair of data belongs to the same category and provide it to the collaborative network for use.

[0071] (3)ResNet-50 module.

[0072] like Figure 5 As shown in Figure 1, as the basic component of the collaborative deep learning network, it includes two complete ResNet-50s. Since ResNet-50 has excellent performance and considering the limited medical data samples, we use pre-trained ResNet-50 as the initial model of the two ResNet-50 modules in the network, denoted as ResNet50-A and ResNet50-B respectively. The pre-trained network includes 50 learning layers, and the learnable parameters of ResNet50-A and ResNet50-B are marked as θ respectively. A and θ B, and do not share parameters with each other. Using the feature sequence X = {x (1) ,x (2) ,…,x (N)} and its corresponding label sequence Y = {y (1) ,y (2) ,…,y (N)} to train each ResNet-50 module, aiming to find a set of parameters θ to reduce the subsequent cross entropy loss:

[0073]

[0074] Where: k is the number of categories, N is the number of training samples in each batch, Z (i) =F(x (i) ,θ) represents the forward operation. This optimization problem can be solved by using a mini-batch stochastic gradient descent algorithm. For k classification problems, all fully connected layers in the original network need to be removed and replaced with a fully connected layer of 1024 neurons and a fully connected layer of k neurons. The parameters of these newly added layers are initialized by a uniform distribution U(-0.05,0.05).

[0075] (4) Collaborative network module.

[0076] To further supervise the training of each ResNet-50 component through the co-labeling of each data pair, we design a co-network consisting of an embedding layer, a fully connected learning layer, and an output layer, as Figure 6 As shown. A pair of data (X A ,X B ) is input into two ResNet-50 components (ResNet50-A, ResNet50-B). The output of the penultimate fully connected layer in ResNet50 is defined as the data depth feature learned by ResNet50, which can be obtained by forward calculation in the following form:

[0077] f A =F(x A ,θ (i) )

[0078] f B =F(x B ,θ (j) )

[0079] Then, the deep features of a pair of data are concatenated as f A,B And input into the collaborative network, the corresponding expected output is the collaborative label of the data pair, defined as follows:

[0080]

[0081] Where: ys =1 means that the input data pair is of the same category and is a positive data pair; s = 0 means that the input data pair is of different categories and is a negative data pair. The collaborative signal can be easily monitored by adding another sigmoid layer and using the following binary cross entropy loss:

[0082]

[0083] Where: θ S are the parameters of the collaborative network, It is the forward computation of the collaborative network, which verifies whether the input data pairs belong to the same category and provides correction feedback if there is a collaborative error.

[0084] The process of the collaborative network algorithm is as follows:

[0085]

[0086] (5) Training process

[0087] The collaborative deep learning model of the present invention consists of two ResNet-50 components and a collaborative network. During end-to-end training, the parameters of each ResNet-50 component and the collaborative network can be updated as follows:

[0088] θ (A) (t+1)=θ (A) ·(t)-η(t)·Δ (A)

[0089] θ (B) (t+1)=θ (B) ·(t)-η(t)·Δ (B)

[0090] θ S(A,B) (t+1)=θ S(A,B) ·(t)-η(t)·Δ S(A,B)

[0091] Where: η(t) is the variable learning rate, S(A,B) represents the collaborative network between ResNet50-A and ResNet50-B.

[0092]

[0093]

[0094]

[0095] Where: λ represents the trade-off between classification error and collaborative error.

[0096] (6)Testing process.

[0097] When the trained collaborative network is used to classify test data, each ResNet-50 part will produce a prediction vector P(A) = (P(A)1, P(A)2, ... P(A)) M ), P(B)=(P(B)1,P(B)2,…P(B) M ), which is the activation value of its last fully connected layer. At this point, the additional collaborative learning network is discarded in the final classification prediction, and the output label of the test data can be predicted as the result y:

[0098]

[0099] The above description of the embodiments is to facilitate the understanding and application of the present invention by those skilled in the art. It is obvious that those skilled in the art can easily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without creative work. Therefore, the present invention is not limited to the above embodiments. Improvements and modifications made by those skilled in the art to the present invention based on the disclosure of the present invention should be within the protection scope of the present invention.

Claims

1. A lung disease classification detection method based on collaborative deep learning and lung auscultation sounds, comprising the following steps: (1) obtaining a certain number of lung auscultation sounds with classification labels, and pre-emphasizing the data signals of the lung auscultation sounds to remove low-frequency noise in the signals; (2) Framing and windowing the pre-emphasized data signal, and using a Hamming window for weighted framing to achieve signal framing. When moving the window, only 1 / 2 or 1 / 3 of the window length is moved, so that there is an overlap between two adjacent frames. (3) Performing a fast Fourier transform on each frame of data signal to obtain the energy distribution of the signal in the frequency spectrum; (4) Extract features from the spectrum of each frame of data signal, take the feature matrix of the entire data signal and its corresponding classification label as a group of samples, and then divide all samples into a training set and a test set; the specific implementation method is: first, pass the spectrum of the data signal through a set of triangular filters to obtain the corresponding Mel frequency features, then use the Mel frequency features to calculate the logarithmic expression of the energy of each frame of data signal, and then perform discrete cosine transform on the logarithmic energy of each frame of data signal to obtain the corresponding multi-order Mel feature cepstrum value to complete feature extraction; finally, combine the Mel feature cepstrum values ​​of the entire data signal into a feature matrix and perform channel amplification on the feature matrix to match the input requirements of the three channels of the model; (5) Construct a classification detection model consisting of two residual neural networks ResNet-50 and a collaborative network. The outputs of the two residual neural networks ResNet-50 are used as the input of the collaborative network. The collaborative network is composed of an embedding layer, a fully connected learning layer and an output layer connected in sequence. The deep features generated by the two residual neural networks ResNet-50 in the second to last fully connected layer are represented as f A and f B , f A With f B Connect to get a combined feature f A,B The output is input into the collaborative network, and the output is used to determine whether the sample pairs input into the two residual neural networks ResNet-50 belong to the same category. If there is a collaborative error, correction feedback is provided; (6) The training set samples are input into two residual neural networks ResNet-50 in pairs. The model is trained using a mini-batch stochastic gradient descent algorithm based on the loss functions of the residual neural network ResNet-50 and the collaborative network to optimize the model parameters. The loss function expression of the residual neural network ResNet-50 is as follows: Where: L(θ) is the loss function of the residual neural network ResNet-50 under parameter θ, N is the number of samples input for each batch, K is the number of categories, is the deep feature generated by the second-to-last fully connected layer of the residual neural network ResNet-50 when the i-th sample in this batch is input, where the subscript j represents the category of the sample input to the residual neural network ResNet-50 detection output, l{y(i)=j} is the discriminant function, if y(i)=j holds, then l{y(i)=j}=1, otherwise l{y(i)=j}=0, y(i) is the true value category of the i-th sample; The loss function expression of the collaborative network is as follows: Where: L s (θ s ) is the collaborative network with parameter θ s The loss function under is the output of the collaborative network, y s is the label value of the collaborative network, y A and B is the true value category of the sample pair input to the two residual neural networks ResNet-50; Finally, the trained model is used to classify and detect the test set samples. Specifically, the test set samples are input into two residual neural networks ResNet-50 in pairs, and two corresponding prediction vectors P(A) and P(B) are obtained, which are the outputs of the last fully connected layer in the residual neural network ResNet-50. At this time, the additional collaborative network is discarded in the final classification prediction, and the final output prediction result is y: in: M is the dimension of the prediction vector, P(A) i and P(B) i are the i-th element values ​​in the prediction vectors P(A) and P(B) respectively.

2. The lung disease classification detection method according to claim 1, characterized in that: In the step (1), a high-pass filter is used to filter the data signal to remove low-frequency noise in the signal and increase the proportion of the high-frequency part, thereby eliminating the inhibitory effect of the sound system on the high-frequency part and highlighting the high-frequency resonance peak.