MW-Net: A Deep Framework for Segmentation and Detection of S1 and S2 Heart Sounds

This paper proposes a Mish W-Net deep framework for precise segmentation of the primary heart sounds, S1 and S2, from PCG signals. The proposed deep framework incorporates a synchro-squeezed wavelet transform for signal-to-image conversion. Additionally, a random search-optimized complex diffusion unsharp masking is proposed to mitigate the issues of noise and image quality after the signal-to-noise conversion. Besides this, to achieve self-regularizing feedback learning, the Mish activation function is employed at the architectural level of the proposed Mish W-Net. The effectiveness of the proposed deep framework is evaluated on the publicly available PhysioNet/Computing in Cardiology PCG dataset. Comparative qualitative and quantitative assessment outperforms the existing methodologies. The proposed method achieves accuracy, precision, recall, F1-score, and intersection over union of 98.28%, 98.77%, 99.50%, 99.13%, and 98.28%. It shows an effective framework that offers precise segmentation and detection of heart sounds.