Semi-Supervised Pulmonary Auscultation Analysis with Cross Pseudo Supervision

Pulmonary auscultation analysis is a crucial part of the diagnosis of respiratory illnesses, yet it remains largely a manual task dependent on the skills of individual physicians. Efforts have been made to use machine learning to automate detection of abnormal lung sounds in auscultation analysis, but typical labeled datasets require experts to engage in time-consuming annotation. In contrast, unlabeled data is relatively easy to obtain. In this paper, we use cross pseudo supervision to leverage a small amount of labeled audio data and a larger amount of unlabeled data to perform automated auscultation analysis. We show that this method significantly outperforms fully supervised models trained only on the labeled data.