Lung sounds classification using convolutional neural networks.

Area of Science:

• Medical acoustics
• Computational intelligence
• Biomedical signal processing

Background:

• Lung sound auscultation is crucial for diagnosing pulmonary disorders but is limited by subjectivity and requires expertise.
• Non-stationary nature of lung sounds complicates accurate analysis, recognition, and distinction.

Purpose of the Study:

• To compare the effectiveness of machine learning approaches for automated lung sound classification.
• To evaluate handcrafted features versus deep learning methods for pulmonary disorder identification.

Main Methods:

• Compared three machine learning approaches: handcrafted features (MFCCs, LBP) with SVM, k-NN, GMM classifiers, and Convolutional Neural Networks (CNNs).
• Utilized MFCC coefficients and Local Binary Patterns (LBP) from spectrograms, with techniques like normalization and whitening.
• Experimented with dataset augmentation on spectrograms to boost CNN performance.

Main Results:

• Convolutional Neural Networks (CNNs) demonstrated superior performance over classifiers using handcrafted features.
• The study evaluated seven lung sound classes: normal, coarse crackle, fine crackle, monophonic wheeze, polyphonic wheeze, squawk, and stridor.

Conclusions:

• CNNs offer a more accurate and robust method for lung sound classification compared to traditional feature extraction techniques.
• Automated systems using CNNs can potentially overcome the limitations of manual auscultation, leading to improved diagnostic accuracy for pulmonary conditions.