Improving The ResNet-based Respiratory Sound Classification Systems With Focal Loss
Jun Li, Xiao Wang, Xiaoqin Wang, Shushan Qiao, Yumei Zhou
Abstract
Automated respiratory sound classification aims to provide a rapid and reliable diagnosis of respiratory disease. However, the database used to develop a lung sounds classification system usually suffers from class imbalance issues, resulting in a lower recall of adventitious sounds compared to the normal class. In this paper, we adopt the focal loss for the ResNet-based systems to solve class imbalance issues. The experiments are conducted on the SPRSound dataset. At the event level, with ResNet18 of 11.3M parameters as the backbone, when trained with class-balanced methods, the best binary and multi-class scores of the validation set are 0.933 and 0.879, and those of the test set are 0.889 and 0.82. While at the record level, contributing to the focal loss, the best ternary score of the validation set is 0.833, achieved by TC-ResNet of 12.2M parameters, and the multi-class score is 0.673 with ResNet18. The ternary and multi-class scores of the test set are 0.718 and 0.533 with TC-ResNet.