Core Concepts
A novel multi-task learning approach is proposed to simultaneously classify lung sounds and lung diseases using deep learning models like 2D CNN, ResNet50, MobileNet, and DenseNet. The MobileNet model achieved the highest accuracy of 74% for lung sound classification and 91% for lung disease classification.
Abstract
The study proposes a multi-task learning (MTL) approach to simultaneously classify lung sounds and lung diseases using deep learning models. Four different deep learning architectures - 2D CNN, ResNet50, MobileNet, and DenseNet - were trained and evaluated on the ICBHI 2017 Respiratory Sound Database.
Key highlights:
- Mel Frequency Cepstral Coefficients (MFCC) were used to extract features from the lung sound recordings.
- The MTL approach allowed for joint optimization of the models to classify both lung sounds and lung diseases, leading to improved performance and reduced training time compared to training separate models.
- Among the four models, the MobileNet architecture achieved the highest accuracy, with 74% for lung sound classification and 91% for lung disease classification.
- The study also investigated risk level prediction for Chronic Obstructive Pulmonary Disease (COPD) using demographic data and machine learning algorithms. The Random Forest classifier provided the best performance with 92% accuracy.
- The proposed framework can aid physicians in efficiently diagnosing lung conditions and communicating the potential causes or outcomes to patients.
Stats
The ICBHI 2017 Respiratory Sound Database was used, which contains 920 audio recordings of lung sounds with annotations.
Quotes
"The MTL for MobileNet model performed better than the other models considered, with an accuracy of 74% for lung sound analysis and 91% for lung diseases classification."
"Results of the experimentation demonstrate the efficacy of our approach in classifying both lung sounds and lung diseases concurrently."