Data augmentation using Variational Autoencoders for improvement of respiratory disease classification

Area of Science:

• Medical informatics
• Artificial intelligence in healthcare
• Respiratory medicine

Background:

• Computerized auscultation of lung sounds is crucial for diagnosing respiratory diseases, but existing datasets like ICBHI are imbalanced.
• Imbalanced datasets hinder the generalization and reliability of deep learning models for lung sound classification.
• Traditional diagnostic methods have limitations that advanced computational approaches aim to surpass.

Purpose of the Study:

• To synthesize respiratory sounds using various Variational Autoencoder (VAE) models, including Multilayer Perceptron VAE (MLP-VAE), Convolutional VAE (CVAE), and Conditional CVAE.
• To evaluate the impact of augmenting an imbalanced respiratory sound dataset with synthesized data on lung sound classification model performance.
• To compare the quality of synthesized respiratory sounds using metrics like Fréchet Audio Distance (FAD).

Main Methods:

• Respiratory sound synthesis using MLP-VAE, CVAE, and Conditional CVAE.
• Dataset augmentation with synthesized respiratory sounds.
• Performance evaluation of lung sound classification models on augmented datasets.
• Quality assessment of synthetic sounds using Fréchet Audio Distance (FAD), Cross-Correlation, and Mel Cepstral Distortion.

Main Results:

• Convolutional VAE (CVAE) and Conditional CVAE demonstrated superior synthetic sound quality with average FAD scores of 11.58 and 11.64, respectively, compared to MLP-VAE's 12.42.
• Augmenting the imbalanced dataset with synthesized sounds led to significant performance improvements in classification metrics for minority classes.
• Marginal performance gains were observed for other classes after data augmentation.

Conclusions:

• Deep learning models show promise for lung sound classification, offering advantages over traditional methods.
• Synthesizing respiratory sounds with VAEs and augmenting imbalanced datasets can significantly enhance the performance of lung sound classification models.
• The study highlights the potential of generative AI techniques in improving the accuracy and reliability of computational diagnostics for respiratory conditions.