Font Size: a A A

Heart Sounds Collection And The Abnormal Detection Based On Convolutional Neural Network

Posted on:2021-01-21Degree:MasterType:Thesis
Country:ChinaCandidate:S ZhangFull Text:PDF
GTID:2370330611992000Subject:Biomedical Engineering
Abstract/Summary:PDF Full Text Request
Objective: Cardiovascular disease is one of the threats to people's health,causing approximately 17.7 million deaths each year,accounting for 31% of global deaths.Heart sound auscultation is non-invasive,convenient and cheap,which makes it play an irreplaceable role in daily clinical diagnosis.However,the current digital stethoscopes still cannot deal with various noises in the process of collecting heart sounds,and the intelligent screening method for abnormal heart sounds still does not achieve good results.Based on the above-mentioned shortcomings,the first part of this article explains the design of a low-noise stethoscope.The second part expounds the abnormal heart sound detection algorithm based on convolutional neural network.It improves the clinical application of heart sound from the perspective of hardware design and algorithm improvement.Methods: In the first part of this article,the auscultation head is designed to replace the traditional "bell-shaped" structure with a pasteable membrane structure,which effectively removes environmental noise and operating noise.At the same time,the white noise of the system is removed using the wavelet threshold method.Transmission distance.We performed diagnostic results testing on the acquisition system and a comparison test with the 3 M ? Littmann? Model 3200 electronic stethoscope for noise immunity and wireless transmission performance,and used the normalized standard deviation(STD)to quantitatively compare the results;In the second part,this paper proposes a novel algorithm structure,which is divided into a pre-training process,a data filter,and a main training process to improve the training stability and accuracy of automatic abnormal heart sound detection.Pre-training and data filters are designed to learn the features that are best suited for classification by removing overlapping data in the feature space,and "cleansing" the training dataset,while the main training is designed to retrain the model on the processed dataset Output the final result.VGG-like Convolutional Neural Network is used in pre-training and main training,and t-SNE and multiple clip k nearest neighbors are used in the data filter.Results: 1.Compared with the 3M stethoscope in the first part of the experiment,the estimated average noise ratio of the acquisition system is 21.26%,and the lowest noise ratio is only 12.47%.In addition,based on the heart sounds recorded by the acquisition system,the diagnosis of all other heart sounds is correct except for two uncertain heart sounds.2.Compared with the latest algorithm,the modified accuracy(Macc)of the report is increased by 5.74%,from 83.35% to 87.99%.The average specificity increased from 73.98% to 90.39%.In addition,the data filter can reduce the Macc standard deviation from 16.63% to 2.08%.Conclusions: 1.Under the condition of ensuring that the diagnostic information is retained,the noise reduction capability is superior to the most advanced equipment.2.While the proposed algorithm further improves the classification accuracy,it can better ensure the stability of training convergence.
Keywords/Search Tags:Heart Sound Signal, Wavelet Transform, Neural Network, Multi-edit-k-nearest-neighbor
PDF Full Text Request
Related items