SSC-UNet: UNet with Self-Supervised Contrastive Learning for Phonocardiography Noise Reduction

TL;DR

The paper addresses the challenge of denoising phonocardiography (PCG) signals for congenital heart disease (CHD) diagnosis in noisy real-world environments where clean training data are scarce. It introduces SSC-UNet, a four-level LU-Net with BiLSTM skip connections implemented in a self-supervised framework that combines Noise2Noise-based denoising with a contrastive learning objective to preserve pathological PCG features. Key contributions include achieving robust denoising without clean data, reducing hallucinations through contrastive learning, and improving downstream classification robustness under hospital-like noise, demonstrated by superior SNR improvements compared to baselines. The work has practical significance for enabling noise-robust PCG-based CHD screening and offers a pathway to extend self-supervised denoising to other physiological signals and potential edge-deployed implementations.

Abstract

Congenital Heart Disease (CHD) remains a significant global health concern affecting approximately 1\% of births worldwide. Phonocardiography has emerged as a supplementary tool to diagnose CHD cost-effectively. However, the performance of these diagnostic models highly depends on the quality of the phonocardiography, thus, noise reduction is particularly critical. Supervised UNet effectively improves noise reduction capabilities, but limited clean data hinders its application. The complex time-frequency characteristics of phonocardiography further complicate finding the balance between effectively removing noise and preserving pathological features. In this study, we proposed a self-supervised phonocardiography noise reduction model based on Noise2Noise to enable training without clean data. Augmentation and contrastive learning are applied to enhance its performance. We obtained an average SNR of 12.98 dB after filtering under 10~dB of hospital noise. Classification sensitivity after filtering was improved from 27\% to 88\%, indicating its promising pathological feature retention capabilities in practical noisy environments.

Figures (7)

• Figure 1: LU-Net with projection head.
• Figure 2: Contrastive learning and its loss function.
• Figure 3: Preprocessing and data augmentation.
• Figure 4: Waveforms and spectrograms of a normal PCG. (a) noisy signal, (b) original signal (hospital noise at 5 dB SNR), denoised signals using (c) N2N (d) CN2N, respectively.
• Figure 5: t-SNE of the projection head with contrastive learning.