ASW-Net: Adaptive Spectral Wavelet Network for Accurate Fetal ECG Extraction

Noninvasive fetal ECG (FECG) is of great significance for monitoring fetal health. However, it is challenging to extract FECG signals from the abdominal ECG signal (AECG) due to the complexity of the task: 1) FECG signals are routinely mixed with noise; 2) FECG signals are aliased with maternal ECG signals in the time and frequency domain. To solve such problems, an adaptive spectral wavelet network (ASW-Net) is proposed for FECG extraction, where the adaptive spectral wavelet module, which can improve the computational efficiency by replacing convolution operation with element-wise Hadamard product in the frequency domain, is first developed to extract FECG components with different frequencies; then, the residual attention module is devised to distinguish FECG signals from noise by capturing waveform details; finally, the inverse spectral wavelet module is designed to reconstruct FECG signals from multi-resolution FECG components. Experiments conducted on the benchmarks demonstrate that the proposed ASW-Net outperforms the state-of-the-art methods.

I. Introduction

Worldwide, there are about 2.5 million stillbirths occur annually, especially in less-developed regions [1]. Thus, it is essential to design a computer-aided tool to monitor fetal well-being during pregnancy and at delivery. Electronic fetal monitoring (EFM) devices, one of the effective computer-aided monitoring tools, have been used widely to monitor fetal growth and improve perinatal outcomes [2]. Recently, most of EFM tools [3] can be summarized into five categories, namely fetal phonocardiography (PCG), Doppler ultrasound, cardiotocography (CTG), fetal magnetocardiography (FMCG), direct fetal ECG, and non-invasive fetal ECG (NI-FECG).