Heart sound signal noise detection method based on multi-scale convolutional network

Abstract

The invention discloses a heart sound signal noise detection method based on a multi-scale convolutional network. The method comprises the following steps: collecting heart sound signal information; cutting into segments with fixed length as network input; then, the cut heart sound fragments are preprocessed; evaluating the quality of the heart sound fragment based on the spectral density ratio, and correcting the label of the low-quality fragment; generating a spectrogram; effective features in spectrograms of different scales are extracted through a convolutional neural network of a hierarchical multi-scale structure, and accurate recognition and classification of heart noise are achieved; training the multi-scale convolutional neural network; and according to an output result of the multi-scale convolutional neural network, outputting a diagnosis result of the cardiac murmur. According to the method, the hierarchical multi-scale convolutional neural network is introduced, so that multi-scale features are effectively fused, and the accuracy and robustness of heart sound signal noise detection and clinical result classification are improved. And the quality of the heart sound fragments is evaluated based on the spectral density ratio, and the labels of the low-quality fragments are corrected, so that the classification performance is further improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical signal processing, and in particular to a heart sound signal murmur detection method based on a multi-scale convolutional network. Background Art

[0002] Early screening of cardiovascular disease (CVD) is essential to reduce the risk of worsening heart disease. Traditional heart murmur detection relies on auscultation by medical staff, but the results are not always reliable due to the limitations of auscultation ability and clinical experience. In recent years, computer-assisted analysis has made significant progress in improving the classification of heart murmurs. However, existing phonocardiogram (PCG) classification methods are mainly divided into two categories: machine learning (ML)-based and deep learning (DL)-based methods, both of which have their own limitations. ML-based methods require manual feature extraction, have strong dependence and lack of robustness; while deep learning-based methods can automatically extract features, but require large-scale data sets to support, and traditional deep convolutional neural networks (CNNs) are difficult to handle long-term and short-term dependencies in heart sound signals. Summary of the invention

[0003] In view of the shortcomings of the prior art, the present invention provides a heart sound signal murmur detection method based on a multi-scale convolutional network, which overcomes the shortcomings of the prior art and has a reasonable design. By introducing a hierarchical multi-scale convolutional neural network, it effectively integrates multi-scale features, improves the accuracy and robustness of heart sound signal murmur detection and clinical result classification. The quality of heart sound segments is evaluated based on the spectral density ratio, and the labels of low-quality segments are corrected, which reduces the misleading of classification results by low-quality records and further improves the classification performance.

[0004] To achieve the above objectives, the present invention is implemented through the following technical solutions: A method for detecting heart sound signal murmur based on a multi-scale convolutional network comprises the following steps: Step S1: Heart sound signal collection: collecting heart sound signal information of multiple patients in the database; Step S2: extracting segments: cutting the collected heart sound signals into segments of fixed length as network input; Step S3: preprocessing: preprocessing the cropped heart sound segments; Step S4: label correction: assess the quality of the heart sound segments based on the spectral density ratio, and correct the labels of low-quality segments; Step S5: generating a spectrogram: converting the preprocessed heart sound signal into a spectrogram using short-time Fourier transform; Step S6: multi-scale convolutional neural network design: design a convolutional neural network with a hierarchical multi-scale structure, extract effective features from spectrograms of different scales through convolutional layers of different depths, and realize accurate recognition and classification of heart murmurs; Step S7: Model training and optimization: Use a large amount of labeled cardiac audio spectrogram data to train the multi-scale convolutional neural network and optimize the network parameters; Step S8: Diagnosis result output: Output the diagnosis result of the heart murmur according to the output result of the multi-scale convolutional neural network.

[0005] Preferably, in step S3, the extracted heart sound segments are preprocessed in a manner including normalization processing, downsampling processing and filtering processing.

[0006] Preferably, in step S6, the convolutional neural network with a hierarchical multi-scale structure uses a multi-scale spectrogram as input, and extracts time-frequency features of different resolutions through spectrograms of different scales; The convolutional neural network with a hierarchical multi-scale structure has a hierarchical structure, which contains four stages of sub-networks. The sub-networks in each stage process spectrograms of different scales respectively. Features are extracted from different scales respectively, and multi-scale features are fused through convolution blocks and global average pooling layers. Finally, the fragments are classified through a linear layer.

[0007] Preferably, in step S6, for the murmur classification of the heart sound signal, a sliding window method is used to continuously classify each heart sound segment, and the label of the entire heart sound segment is determined by calculating the proportion of frames predicted to be "abnormal".

[0008] Preferably, in step S7, Adamw optimizer is used for training, the maximum training rounds are set to 100, the initial learning rate is 10-3; and when the training loss stops decreasing within five rounds, the learning rate is multiplied by 0.1; and cross entropy with 0.1 label smoothing is used as the loss function.

[0009] The present invention provides a method for detecting murmurs of heart sound signals based on a multi-scale convolutional network. The method has the following beneficial effects: by introducing a hierarchical multi-scale convolutional neural network, by receiving a multi-scale spectrum graph as input, and by constructing long-term and short-term dependencies between multi-scale features, the multi-scale features are effectively integrated, thereby improving the accuracy and robustness of heart sound signal murmur detection and clinical result classification. By evaluating the quality of heart sound segments based on the spectral density ratio and correcting the labels of low-quality segments, the misleading effect of low-quality records on classification results is reduced, thereby further improving the classification performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the present invention or the technical solutions in the prior art, the drawings required for describing the prior art are briefly introduced below.

[0011] Figure 1 A simplified flowchart of the steps of the center tone signal noise detection method of the present invention; Figure 2 This is a conceptual diagram of the central tone detection algorithm of the present invention; Figure 3 Schematic diagram of the multi-scale convolutional neural network structure in the present invention; Figure 4 The best confusion matrix diagram for patient classification in the experimental example of the present invention. DETAILED DESCRIPTION

[0012] In order to make the purpose, technical solutions and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the present invention. Example

[0013] like Figure 1-3 As shown, a method for detecting heart sound signal murmur based on a multi-scale convolutional network comprises the following steps: Step S1: Heart sound signal collection: collect heart sound signal information of multiple patients in the database; in this embodiment, the database used is the publicly released data set of PhysioNet Challenge 2022, which contains 3163 PCG records from 942 patients; Step S2: Extracting segments: In order to speed up data loading without losing noise information, the collected heart sound signals are cut into segments of fixed length as network input; specifically, the recording sampling frequency is downsampled from 4000 Hz to 2000 Hz.

[0014] Step S3: Preprocessing: Preprocess the cropped heart sound segments, including normalization, downsampling and filtering. Normalization is used to eliminate the difference in signal amplitude between different acquisition devices; downsampling is used to reduce the amount of calculation and increase the processing speed; filtering is used to remove noise and interference components in the signal.

[0015] Step S4: Label correction: Since the heart sound segment contains many low-quality segments due to environmental noise, artifacts, body friction, etc., these segments will mislead the optimization of CNN. Therefore, it is necessary to evaluate the quality of the murmur label of the segment for correction. Since the frequency of normal heart sounds is between 20-200 Hz, the energy of most murmurs is much lower than that of heart sounds. Therefore, the quality of heart sound segments can be evaluated based on the spectral density ratio, and the labels of low-quality segments can be corrected; specifically, the selected evaluation criterion is the spectral density ratio between the 20-200 Hz frequency band and the full frequency band (0-1000 Hz); if the spectral density ratio is greater than 30%, the murmur label of the segment follows the recorded label; otherwise, it is relabeled as "unknown"; this label correction is only for the murmur label, while the result label remains unchanged.

[0016] Step S5: Spectrogram generation: The pre-processed heart sound signal is converted into a spectrogram using short-time Fourier transform; the spectrogram can intuitively display the distribution of the heart sound signal at different frequencies and times, providing a basis for subsequent image recognition.

[0017] Step S6: multi-scale convolutional neural network design: design a convolutional neural network with a hierarchical multi-scale structure, use a multi-scale spectrogram as input, extract time-frequency features of different resolutions in different scale spectrograms through convolutional layers of different depths, and realize accurate recognition and classification of heart murmurs; Specifically, Figure 3 The figure shows the overall structure of the hierarchical multi-scale convolutional neural network, which first receives three segments of spectrograms as input and outputs three types of murmur predictions. For the outcome prediction, the patient information is combined and the classification result is output. The convolutional neural network with a hierarchical multi-scale structure has a hierarchical structure, including four stages of sub-networks, each stage containing layers of different depths for processing spectrograms of different scales; then extract features from different scales respectively, for example, in the first stage, two sub-networks are used to convolve features from scale 1 and scale 2. The features of these two scales are then concatenated in the channel dimension and passed to the next stage. The fourth stage aggregates the multi-scale convolutional features through global average pooling, thereby fusing the multi-scale features through convolution blocks and global average pooling layers, and finally classifying the segments through a linear layer.

[0018] For patient classification, patient information including age, sex (one-hot encoding), pregnancy status, height, and weight are used as additional information to distinguish patients with abnormal clinical outcomes. Figure 2 As shown in Figure 1, patient features are extracted from this information through a four-layer multilayer perceptron. The final result prediction combines convolutional features and patient features.

[0019] As for the murmur classification of heart sound signals, in order to achieve continuous classification of the entire heart sound record, the present invention adopts a sliding window method. Specifically, the heart sound record is cut into segments of fixed length (such as 3 seconds), and each segment is continuously classified using a sliding window (such as a window width of 3 seconds and a step length of 1 second). For each frame that is slid over by multiple windows, its label is calculated by the average distribution probability of the window. If the number of frames predicted to be "abnormal" exceeds 1 / 3 of the total number of frames, the record is predicted to be "abnormal". By adopting the sliding window method, the heart sound signal can be analyzed more carefully and continuously, thereby improving the accuracy and reliability of murmur detection.

[0020] Specifically, if the predicted unknown label accounts for more than 80% of the sequence labels, the record is classified as unknown. Otherwise, the record is classified based on the majority of sequence labels (excluding unknown). For patient classification, if there is a position record classified as present, the patient is marked as present. For the absent and unknown cases, the patient is classified based on the majority of position record labels. When the number of absent and unknown records for a patient is the same, absent takes precedence.

[0021] Step S7: Model training and optimization: Use a large amount of labeled cardiac audio spectrogram data to train the multi-scale convolutional neural network and optimize the network parameters; Specifically, the Adamw optimizer was used for training, the maximum training rounds were set to 100, the initial learning rate was 10-3; and when the training loss stopped decreasing within five rounds, the learning rate was multiplied by 0.1; and the cross entropy with 0.1 label smoothing was used as the loss function.

[0022] Step S8: Diagnosis result output: Output the diagnosis result of the heart murmur according to the output result of the multi-scale convolutional neural network.

[0023] Specifically, the result classification strategy is similar to the murmur classification. The sequence labels per second are also obtained through the sliding window. If the number of frames predicted as "abnormal" exceeds 1 / 3, the record is predicted as "abnormal". When a patient has at least one record predicted as "abnormal", our strategy diagnoses the patient as "abnormal". Otherwise, the patient is diagnosed as "normal".

[0024] Experimental example This embodiment evaluates the above method by using a five-fold cross validation based on patients. The scoring indicators for the murmur weighted accuracy and result challenge score are shown in the figure.

[0025] The heart sound signal murmur detection method based on multi-scale convolutional network of the present invention achieved an average murmur classification accuracy of 91.37% on the five-fold cross-validation segment. In the patient-level murmur classification, the average accuracy was 83.78%, and the average weighted accuracy of murmur classification reached 0.81. Figure 4 The best confusion matrix for patient classification is shown. The overall accuracy was 89.94%, with 85.0% accuracy for "present", 53.84% for "unknown", and 94.85% for "absent". The weighted murmur accuracy was 0.853. In terms of patient outcome classification, our method achieved an average accuracy of 56.83% and an average sensitivity of 84.33%. In the blind validation set, the algorithm achieved a weighted murmur accuracy of 0.806.

[0026] The present invention introduces a hierarchical multi-scale convolutional neural network, receives a multi-scale spectrogram as input, and constructs long-term and short-term dependencies between multi-scale features to effectively integrate multi-scale features, thereby improving the accuracy and robustness of heart sound signal murmur detection and clinical result classification. By evaluating the quality of heart sound segments based on the spectral density ratio and correcting the labels of low-quality segments, the misleading effect of low-quality records on classification results is reduced, further improving the classification performance.

[0027] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A heart sound signal murmur detection method based on a multi-scale convolutional network, characterized in that: The following steps are involved: Step S1: Heart sound signal collection: collecting heart sound signal information of multiple patients in the database; Step S2: extracting segments: cutting the collected heart sound signals into heart sound segments of fixed length as network input; Step S3: preprocessing: preprocessing the cropped heart sound segments; Step S4: label correction: assess the quality of the heart sound segments based on the spectral density ratio, and correct the labels of low-quality segments; Step S5: generating a spectrogram: converting the preprocessed heart sound signal into a spectrogram using short-time Fourier transform; Step S6: multi-scale convolutional neural network design: design a convolutional neural network with a hierarchical multi-scale structure, extract effective features from spectrograms of different scales through convolutional layers of different depths, and realize accurate recognition and classification of heart murmurs; Step S7: Model training and optimization: Use a large amount of labeled cardiac audio spectrogram data to train the multi-scale convolutional neural network and optimize the network parameters; Step S8: Diagnosis result output: Output the diagnosis result of the heart murmur according to the output result of the multi-scale convolutional neural network.

2. The method for detecting heart sound signals based on a multi-scale convolutional network according to claim 1, characterized in that: In step S3, the extracted heart sound segments are preprocessed in a manner including normalization, downsampling and filtering.

3. The method for detecting heart sound signals based on a multi-scale convolutional network according to claim 1, characterized in that: In step S6, the convolutional neural network with a hierarchical multi-scale structure uses a multi-scale spectrogram as input, and extracts time-frequency features of different resolutions through spectrograms of different scales; The convolutional neural network with a hierarchical multi-scale structure has a hierarchical structure, which contains four stages of sub-networks. The sub-networks in each stage process spectrograms of different scales respectively. Features are extracted from different scales respectively, and multi-scale features are fused through convolution blocks and global average pooling layers. Finally, the fragments are classified through a linear layer.

4. The method for detecting heart sound signals based on a multi-scale convolutional network according to claim 1, characterized in that: In step S6, for the noise classification of the heart sound signal, a sliding window method is used to continuously classify each heart sound segment, and the label of the entire heart sound segment is determined by calculating the proportion of frames predicted to be "abnormal".

5. The method for detecting heart sound signals based on a multi-scale convolutional network according to claim 1, characterized in that: In step S7, Adamw optimizer is used for training, the maximum training rounds are set to 100, and the initial learning rate is 10-3; and when the training loss stops decreasing within five rounds, the learning rate is multiplied by 0.1; and the cross entropy with 0.1 label smoothing is used as the loss function.

Images