Enhancing Lung Sound Classification Using Transfer Learning with ResNet50 and Mel Spectrogram Pre-processing

Abstract

Lung disease is one of the leading causes of illness and death worldwide. Human lung sounds from breathing can vary from normal to abnormal. The presence of crackles, wheezes, rhonchi, and a combination of crackles-wheezes can indicate abnormal lung sounds. This study presents transfer learning with ResNet50, which can explore and prove its effectiveness in handling the task of classifying lung sound signals, including complex classes such as wheezing crackles. Performance measurements are made using complete evaluation metrics, and an in-depth analysis of model performance is conducted, including accuracy, precision, recall, F1 score, AUC, and confusion matrix. Providing a more comprehensive understanding of the model's ability to classify lung sound signals. The results are compared using various optimizers, such as Adam, SGD, Adamax, and RMS Sprop. The research stages include dataset collection, pre-processing, data augmentation, fold cross-validation, and transfer learning with ResNet50. The model shows superior performance when the right optimizer, such as Adamax, is selected. Performance measurements obtained an accuracy value of 95.33%, precision of 95.46%, recall of 93.33%, F1 score of 95.24%, and AUC of 99.63%. The results of this study can provide important insights for the medical community.