ECGDenoiser: A Magnitude-Aware Deep Learning Framework with Phase Retrieval for Electrocardiogram Signal Enhancement

High-quality electrocardiogram (ECG) signals play a pivotal role in many heart-related healthcare applications. However, various types of real-world noises can contaminate the ECG signal during acquisition, thereby impairing its practical applicability. Thus, it is of great significance to develop an effective method for noise filtering in contaminated ECG signals. Most existing techniques predominantly focus on reducing noise in contaminated ECG signals in time domain, overlooking the potential benefits of time-frequency analysis. In this study, we propose a magnitude-aware deep learning framework with phase retrieval for denoising contaminated ECG signals (ECGDenoiser) based on time-frequency domain analysis. Firstly, the corrupted magnitude spectrum and phase spectrum are obtained by applying the short-time Fourier transform (STFT) to the contaminated ECG signal. Secondly, the ConvNeXtV2 Attention UNet is proposed to separate noise magnitude spectrum from the corrupted spectrum, and a fast Griffin-Lim algorithm is introduced to reconstruct the phase spectrum. Finally, the inverse STFT is adopted to recover the signal through estimated ECG magnitude spectrum and enhanced phase spectrum. The denoising performance of the proposed method is denoted by root mean square error (RMSE) and signal-to-noise ratio (SNR). Experiments conducted on single-channel and multi-channel ECG datasets demonstrate that ECGDenoiser surpasses conventional filtering techniques and deep learning-based methods in addressing seven distinct categories of noise interference. The code of this study is released at https://github.com/Shurun-Wang/ECGDenoiser.