Lightweight Skip Connections With Efficient Feature Stacking for Respiratory Sound Classification

Abstract

As the number of deaths from respiratory diseases due to COVID-19 and infectious diseases increases, early diagnosis is necessary. In general, the diagnosis of diseases is based on imaging devices (e.g., computed tomography and magnetic resonance imaging) as well as the patient's underlying disease information. However, these examinations are time-consuming, incur considerable costs, and in a situation like the ongoing pandemic, face-to-face examinations are difficult to conduct. Therefore, we propose a lung disease classification model based on deep learning using non-contact auscultation. In this study, two respiratory specialists collected normal respiratory sounds and five types of abnormal sounds associated with lung disease, including those associated with four lung lesions in the left and right anterior chest and left and right posterior chest. For preprocessing and feature extraction, the noise was removed using three pass filters (low, band, and high), and respiratory sound features were extracted using the Log-Mel Spectrogram-Mel Frequency Cepstral Coefficient followed by feature stacking. Then, we propose a lung disease classification model of dense lightweight convolutional neural network-bidirectional gated recurrent unit skip connections using depthwise separable convolution based on the extracted respiratory sound information. The performance of the classification model was compared with both the baseline and the lightweight models. The results indicate that the proposed model achieves high performance and has an accuracy of 92.3%, sensitivity of 92.1%, specificity of 98.5%, and f1-score of 91.9%. Using the proposed model, we aim to contribute to the early detection of diseases during the COVID-19 pandemic.

Interpretation of lung disease classification with light attention connected module

“…We compared the model to a previous study using the same Littmann 3200 data set. Choi et al [23] emphasized the respiratory sound information using feature stacking, but this study emphasized the features of feature information using LACM based on ECA-Net. As shown in Fig.…”

“…

Fig. 8 Comparison of attention models: baseline, No-attention, CBAM, SENet, proposed model, and [23] .

…”

“…The energy coefficient of the characteristic frequency section was obtained by performing a discrete cosine transform (DCT) on the log-Mel spectrogram. Therefore, in this study, the log-Mel spectrogram MFCC [23] was used as an input for the proposed model. The MFCC generation process shown in Fig.…”

“…A depthwise separable convolution was used to improve the model with deeper layers and reduce the amount of computation. The structure is a convolution constructed by merging a depthwise structure that separates the feature map for each channel and performs an operation and a pointwise convolution structure that merges multiple channels into one new channel [23] .…”

“…They attained the highest COPD identification performance by using the proposed deep belief networks. Choi et al [23] proposed a lightweight CNN-GRU skip connection combining a CNN and BiGRU by collecting respiratory sounds from outpatients and inpatients using Littmann 3200. Through feature stacking, the characteristics of respiratory sounds were emphasized, and a high performance of 92.3 % was confirmed.…”