이중 헤드 시간 영역 분해를 활용한 심장-폐음 분리 개선을 위한 합성 혼합물 감독 프레임워크

초록

Heart-Lung sound separation is a key preprocessing task for AI-assisted computerized auscultation, particularly because cardiac and respiratory sounds are often superimposed in single-channel recordings. To address the limited availability of paired ground-truth cardiopulmonary data, this study proposes DHD-Net, a dual-head time-domain decomposition network trained through synthetic mixture supervision. DHD-Net integrates a shared encoder, two source-specific decoder branches, band-pass filtering, temporal shifting, gain perturbation, and joint waveform-spectral learning. The proposed model was evaluated against four baselines: mixed input, band-pass filtering, NMF-based separation, and Wave-U-Net. In paired evaluation, DHD-Net achieved the highest mean SI-SDR improvement among the compared models, with an overall gain of 0.146 dB. Source-wise analysis showed a -0.775 dB change for heart sounds and a 1.066 dB improvement for lung sounds, indicating stronger reconstruction performance for lung sounds. Although both sources improved simultaneously in only 51.7% of cases and mean Source performance improved in 56.6% of paired samples, statistical and qualitative analyses suggest that DHD-Net has meaningful potential as an interpretable AI framework under limited paired-data conditions.

키워드