Automatic Classification of Respiratory Sounds by Improving the Loss Function of ResNet

Respiratory diseases cause 8 million deaths annually, and this number is expected to increase.Breath auscultation, a primary diagnostic method, is noninvasive, repeatable, and immediate, but faces challenges such as reliance on skilled practitioners, difficulty in quantitative assessment, and limited accessibility in developing regions or disaster sites.To address these issues, we developed a deep learning-based breath sound classification system using the ICBHI 2017 dataset.Our method classifies breath sounds into four categories: Normal, Crackle, Wheeze, and Crackle and Wheeze.We use ResNet-34 as the base model, which is enhanced with CBAM for better spatial and channel feature extraction.To deal with class imbalances, we incorporate Focal Loss.The system achieves Accuracy of 0.732, SE of 0.607, SP of 0.843, and ICBHI Score of 0.725.