2021/04/30 by Truc Nguyen, Franz Pernkopf, Nguyen, Truc +1
Medicine · Computer Science · #Phonocardiography and Auscultation Techniques #Music and Audio Processing #Respiratory and Cough-Related Research
paper · pdf · doi:10.48550/arxiv.2104.14921
Large annotated lung sound databases are publicly available and might be used\nto train algorithms for diagnosis systems. However, it might be a challenge to\ndevelop a well-performing algorithm for small non-public data, which have only\na few subjects and show differences in recording devices and setup. In this\npaper, we use transfer learning to tackle the mismatch of the recording setup.\nThis allows us to transfer knowledge from one dataset to another dataset for\ncrackle detection in lung sounds. In particular, a single input convolutional\nneural network (CNN) model is pre-trained on a source domain using ICBHI 2017,\nthe largest publicly available database of lung sounds. We use log-mel\nspectrogram features of respiratory cycles of lung sounds. The pre-trained\nnetwork is used to build a multi-input CNN model, which shares the same network\narchitecture for respiratory cycles and their corresponding respiratory phases.\nThe multi-input model is then fine-tuned on the target domain of our\nself-collected lung sound database for classifying crackles and normal lung\nsounds. Our experimental results show significant performance improvements of\n9.84% (absolute) in F-score on the target domain using the multi-input CNN\nmodel based on transfer learning for crackle detection in adventitious lung\nsound classification task.\n