Enhancing heart sound signal denoising: unveiling the impact of time-frequency transformation in U-Net performance

Cardiovascular diseases are the leading cause of death worldwide. Automatic heart sound analysis offers a low-cost and highly promising diagnostic method, but its effectiveness is often compromised by noise during recording. Nowadays, most advanced blind source separation methods typically employ deep learning techniques based on a U-Net architecture with a short-time Fourier transform (STFT) for time-frequency representation of audio signals. However, alternative time-frequency representations have not been thoroughly explored. In this paper, we enhance phonocardiogram signal denoising by utilizing four distinct time-frequency representations: STFT, continuous wavelet transform (CWT), wavelet synchrosqueezed transform, and the S-transform, processed through a U-Net. We evaluated our denoising algorithm by contaminating heart sound signals with four different types of noise, including non-stationary real-world nuisance signals, at -5 dB and 0 dB signal-to-noise ratios (SNRs). Our results showed that the CWT achieved the best denoising performance, with gains of over 18 dB for heart sounds contaminated with speech at -5 dB SNR. We found that the CWT significantly improves by 2 to 8 dB over STFT while requiring less than double the computational resources. The method’s effectiveness is demonstrated in the time domain, time-frequency, and power spectral density spaces, resulting in a denoising model that surpasses current state-of-the-art techniques.