Kim June-Woo ; Bae Sangmin ; Cho Won-Yang ; Lee Byungjo ; Jung Ho-Young

Abstract

Despite the remarkable advances in deep learning technology, achievingsatisfactory performance in lung sound classification remains a challenge dueto the scarcity of available data. Moreover, the respiratory sound samples arecollected from a variety of electronic stethoscopes, which could potentiallyintroduce biases into the trained models. When a significant distribution shiftoccurs within the test dataset or in a practical scenario, it can substantiallydecrease the performance. To tackle this issue, we introduce cross-domainadaptation techniques, which transfer the knowledge from a source domain to adistinct target domain. In particular, by considering different stethoscopetypes as individual domains, we propose a novel stethoscope-guided supervisedcontrastive learning approach. This method can mitigate any domain-relateddisparities and thus enables the model to distinguish respiratory sounds of therecording variation of the stethoscope. The experimental results on the ICBHIdataset demonstrate that the proposed methods are effective in reducing thedomain dependency and achieving the ICBHI Score of 61.71%, which is asignificant improvement of 2.16% over the baseline.