Fetal ECG Extraction Based On Attention And Multi-Scale Residual Contraction U-Net

The fetal ECG signal records the potential changes of fetal heart movement.Through the characteristic analysis of its waveform,it can determine whether there are congenital heart disease,fetal intrauterine hypoxia,distress and other symptoms,which ensures the growth and health of the fetus during pregnancy,and is an effective means of perinatal fetal monitoring.Non-invasive method collects fetal ECG signals by placing electrodes in the mother’s abdomen,without harming pregnant women and fetus,and becomes the development direction of fetal ECG signal extraction.However,non-invasive methods need to be transmitted through the abdominal wall of the mother,and the collected ECG signals,in addition to fetal ECG signals,also interfere with maternal ECG components.Therefore,obtaining clean fetal ECG signals is the focus of non-invasive methods.In this thesis,the feature extraction of adaptive fetal ECG signal combined with deep learning is deeply studied,and a U-Net model based on multi-scale residual contraction and attention mechanism is proposed to extract fetal ECG signal.The main research work is as follows:(1)Aiming at the problems of single feature scale in U-Net model extraction,incomplete removal of maternal ECG features,and loss of down-sampling pooled information,a method of fetal ECG signal extraction based on multi-scale residual contraction U-Net model is proposed.Based on U-Net model,multi-scale module Inception is introduced to extract the features of fetal ECG signal at multi-scale;Aiming at the problem that the residual contraction module with different thresholds per channel does not fully utilize the time domain information,the time domain attention is introduced and the original channel attention is combined to form the mixed domain attention for adaptive threshold learning to improve the ability of the model to remove the mother’s ECG characteristics;Finally,Maxpool of U-Net is replaced by Softpool to reduce the information loss in the down-sampling process and retain the small features in the fetal ECG signal.Through the analysis of experimental results,the proposed multi-scale residual contraction U-Net model can effectively remove the mother’s ECG components and extract relatively pure fetal ECG signals.(2)In order to further improve the extraction effect of the multi-scale residual contraction U-Net model at the maternal-fetal ECG coincidence,a multi-scale residual contraction U-Net model based on the attention mechanism is proposed.By comparing the attention method and adjusting the adding position,it can better screen the fetal ECG features,suppress the useless features,and improve the extraction accuracy of the fetal ECG signal at the maternal-fetal ECG coincidence.Through the comparison of experimental results,it is confirmed that the multi-scale residual contraction U-Net-Behind model based on SE attention can effectively improve the extraction accuracy of fetal ECG signals at the maternal-fetal ECG coincidence.