Utilizing Various Transfer Learning Approaches for the Identification of Lung Respiratory Sounds

Abstract

Introduction: The sounds produced by the lungs when breathing might provide important information to physicians. Based on our findings, we recommend a deep learning-based approach to the prediction of breathing-related lung sounds. The Proposed model was trained in lung sounds collected from people suffering from a broad variety of respiratory conditions. The research improves classifying lung sounds, by audio to image spectrogram features is taken and used to train a deep convolutional neural network.

Objectives: The objective of this study is to improve the accuracy with which deep learning can anticipate pulmonary breath noises such as wheezes, crackles, and normal breathing. The harnessing potential of deep learning may develop an accurate and objective method for forecasting unique respiratory lung sounds, which will aid in the early identification and treatment of respiratory disorders.

Methods: The research improves classifying lung sounds, by audio to image spectrogram features is taken and used to train a deep convolutional neural network. The proposed technique accurately predicts many different types of respiratory lung sounds, demonstrating the promise of deep learning in this domain.

Results: Theresults demonstrates that deep learning models are capable of reliably predicting a range of breathing-related lung sounds. The VGG16, ResNet50, and proposed CNN models are examined and contrasted on the Lung Sound Dataset. The findings show that the proposed CNN model has a higher accuracy (95%), precision (95%), recall (92%), and F1-score (93%), which are all indices of the model's ability to predict different types of respiratory lung sounds, than the other two models.

Conclusions: In conclusion, this research results have important implications for the development of automated diagnostic tools that might help doctors make correct diagnoses of respiratory disorders more quickly and accurately.