Robust Denoising of Phonocardiogram Signals Using Time-Frequency Analysis and U-Nets

Cardiovascular diseases are the cause of many deaths worldwide every day. Automated cardiac auscultation is a promising diagnosis method; however, one of its main disadvantages is that it is prone to receiving too much noise during sound recording, which hinders diagnosis. In the available literature, most phonocardiogram (PCG) denoising methods have been evaluated using only synthetic sources such as white noise; this work proposes a more realistic approach. In this paper, the denoising process occurs in the time-frequency domain. More specifically, we compute the Short-Time Fourier Transform (STFT) of the contaminated PCG signal and train a U-Net to recognize normal and pathological cardiac sounds from noise. We are unaware of previous attempts to develop a robust PCG-denoising algorithm capable of simultaneously removing noise signals from four different sources: additive white Gaussian noise (AWGN), additive pink Gaussian noise (APGN), speech, and real PCG background noise. Since we are limited by the relatively small number of clean PCG signals available, we also propose a method for high-quality phonocardiogram data augmentation. In our tests, the proposed method exhibits high performance even in unfavorable scenarios since it can denoise a PCG signal contaminated with a signal-to-noise ratio (SNR) of −5 dB with average improvements ranging from 11.85 dB up to 17.60 dB, depending on the noise type used to degrade the cardiac signal. This method could significantly improve the performance of automatic cardiac sound classification algorithms in noisy environments but could also be used in electronic stethoscopes.