Research On CHD-PAH Aided Diagnosis Algorithm Based On Deep Learning And Fusion Feature

Congenital heart disease(CHD)is a common innate malformation heart disease in young children,It usually occurs when the baby’s heart and blood vessels do not develop properly during embryonic life.Congenital heart disease operation has age limitation,and it is difficult to prevent,so it needs early detection and early treatment.there is a high incidence of congenital heart disease in some remote mountainous areas of the plateau.However,due to the limitation of medical conditions,the lack of medical equipment,as well as the scarcity of doctors and other problems,lead to the patients miss the optimal treatment period.Pulmonary hypertension associated with congenital heart disease is a disease caused by congenital heart disease.Cardiac auscultation is a common method for initial diagnosis of congenital heart disease and pulmonary hypertension.Echocardiography is used to confirm suspected cases after auscultation.Because auscultation needs the doctor to have the professional level and the rich experience,it is easy to cause misdiagnosis and missed diagnosis,This is a pity for missing missing the golden period of surgery in the age of minors.So with the help of machines to assist the doctors to treat the congenital heart disease and pulmonary hypertension is particularly important.The purpose of this study is to diagnose congenital heart disease and pulmonary hypertension by processing and analyzing heart sound signals.The main work of this paper is as follows:1.For the collected heart sound signals containing noise,the pretreatment of heart sound is mainly noise reduction and segmentation,and the denoising method is wavelet transform.A heart sound segmentation method according to Shannon envelope and Viola integral envelope fusion.The segmentation effect is tested experimentally,and the segmentation accuracy is improved.Classification of preprocessed and original heart sounds on the same data set showed an improvement in accuracy of about 8%.2.The pretreated heart sounds were calculated by The Teager-Kaiser energy operator(TKEO),and then the TK signal energy obtained by the calculation was extracted from the spectrum map based on Mel-Frequency Cepstral coefficients(MFCC)to obtain the frequency domain characteristics.The TK-MEL feature of the fusion of the two was used as the feature of subsequent classification.3.The extracted TK-MEL heart sound features were tested by Recurrent neural network(RNN)training,and the classification results were obtained.The classification model was obtained by using normal heart sounds and congenital heart sounds as source tasks,and then the Trada Boost algorithm of transfer learning was used to transfer the classification model to further realize the classification of pulmonary hypertension heart sounds and congenital heart disease.