Heart Sound Intelligent Analysis Based On Deep Convolutional Neural Networks

Cardiovascular disease is one of the major causes of morbidity and mortality.It brings great misfortune and deepness for both family and society,particularly for the children with congenital heart diseases.Timely detection with properly treatment is the best way to improve the situation.As an effective and non-invasive approach,heart sound auscultation can receive the physical condition of the heart simply and quickly,which is a primary diagnosis method for detecting cardiovascular diseases.Nevertheless,auscultation by physician suffers from the subjectivity,and it requires a large number of clinical experience.Because of the objectivity and effectiveness,the computer-aided heart sound analysis has become a promising area for researchers nowadays.Because of insufficient quantity and quality of training data,the previous methods of heart sound analysis are mainly based on conventional handcraft-feature extraction and machine learning.Nonetheless,those methods are difficult to be further improved in practical applications,due to the complexity of the heart sounds and the interference of environmental noises.After the publication of the PhysioNet/CinC 2016 publicly available dataset,some deep learning-based methods have been proposed.However,due to the simple network architecture they used,poor performance with excessive parameter consumption they obtained,which shows limitation in applying to resource-constrained situation such as mobile applications or smart wearable devices.In order to cope with the aforementioned issues,the research of this thesis focus on two different aspects,i.e.,heart sound data and the deep convolutional neural network models for heart sound diagnosis.And the main contents of this thesis are as follows:1.A pediatric heart sound dataset is proposed for training deep learning model in this thesis.By investigating the mechanism and characteristics of heart sound generation,and summarizing the existing publicly available heart sound datasets,some differences between heart sounds from adult and children are found.In order to make deep learning-based pediatric heart sound analysis possible,heart sounds are collected and annotated according to the results of echocardiography.After manually removing the irrelevant environmental noises,the pediatric heart sound dataset with 528 recordings is built.2.The heart sound intelligent analysis based on one-dimensional deep convolutional neural networks is proposed for normal and abnormal heart sound diagnosis.Aim at improving the accuracy and reducing the parameter consumption of current methods,the one-dimensional DenseNet-based and the one-dimensional CliqueNet-based deep convolutional neural network are proposed for heart sound diagnosis,which are from the idea of feature map reuse,attention mechanism and feature map decoupling.Experiments on PhysioNet/CinC 2016 and the self-built pediatric heart sound dataset show that the superiority of two proposed networks in terms of diagnosis accuracy and parameter consumption.In addition,this thesis also attempts to explore the interpretability of the deep learning-based models in the heart sound diagnosis task.