Improved Lung Sound Classification Model Using Combined Residual Attention Network and Vision Transformer for Limited Dataset

According to WHO data, the prevalence of respiratory disorders is increasing, exacerbated by a shortage of skilled medical professionals. Consequently, there is an urgent need for an automated lung sound classification system. Current methods rely on deep learning, but limited lung sound data resulted in low model accuracy. The widely used ICBHI 2017 dataset has an imbalanced class distribution, with a normal class at 52.8%, wheezing at 27.0%, crackles at 12.8%, and combined wheeze and crackles at 7.3%. The imbalance of the dataset may affect the model's efficiency and performance in classifying lung sounds. Given these data limitations, we propose a hybrid model, combining residual attention network (RAN) and vision transformer (ViT), to construct an effective respiratory sound classification model with a small dataset. We employ feature fusion techniques between convolutional neural network (CNN) feature maps and image patches to enrich lung sound features. Additionally, our preprocessing involves bandpass filtering, resampling sounds to 16 kHz, and normalizing volume to 15 dB. Our model achieves impressive ICBHI scores with 97.28% specificity, 92.83% sensitivity, and an average score of 95.05%, marking a 10% improvement over state-of-the-art models in previous research.