RespireNet: A Deep Neural Network for Accurately Detecting Abnormal Lung Sounds in Limited Data Setting

Respiratory diseases like asthma, chronic obstructive pulmonary disease (COPD), lower respiratory tract infection, lung cancer, and tuberculosis are the leading causes of death worldwide [14], constituting four of the 12 most common causes of death. Early diagnosis has been found to be crucial in limiting the spread of respiratory diseases, and their adverse effects on the length and quality of life. Listening to chest sounds using a stethoscope is a standard method for screening and diagnosing lung diseases. It provides a low cost and non-invasive screening methodology, avoiding the exposure risks of radiography and patient-compliance requirements associated with tests such as Spirometry.

There are a few drawbacks of stethoscope-based diagnosis: requirement of a trained medical professional to interpret auscultation signals, and subjectivity in interpretations causing inter-listener variability. These limitations are exacerbated in impoverished settings and during pandemic situations (such as COVID-19), due to shortage of expert medical professionals. Automated analysis of respiratory sounds can alleviate these drawbacks, and also help in enabling tele-medicine applications to monitor patients outside a clinic by less-skilled workforce such as community health workers.

Algorithmic detection of lung diseases from respiratory sounds has been an active area of research [17, 20] especially with the advent of digital stethoscopes. Most of these works focus on detecting abnormal respiratory sounds of wheeze and crackle. Wheeze is a typical symptom of asthma and COPD. It is characterized by a high-pitched continuous sound in the frequency range of 100-2500Hz and duration above 80 msec [3, 19]. Crackles, which are associated with COPD, chronic bronchitis, pneumonia and lung fibrosis [7, 18], have a discontinuous, non-tonal sound, around frequency of $\sim$650 Hz and duration of 5 msec (for fine crackles), or frequency of 100-500 Hz and duration of 15 msec (for coarse crackles).

Although early works focused on hand-crafted features and traditional machine learning [8, 4], more recently, deep learning based methods have received the most attention [9, 1, 12]. For training DNNs, a time-frequency representation of the audio signal such as Mel-spectrograms [25, 10, 1], stacked MFCC features [25, 2, 16, 13, 9] or optimized S-transform spectrogram [6] is used. This 2D “image” is then fed into CNNs [16, 2], RNNs [15, 9], or hybrid CNN-RNNs [1] to learn robust high dimensional representations.

It is well known that DNNs are data hungry and typically require large datasets to achieve good performance. In this work, we use the ICBHI challenge dataset [21], a popular respiratory sound dataset. In spite of being the largest publicly available dataset, it has only 6898 breathing cycle samples, which is quite small for training deep networks. Thus, a big focus of our work has been on developing a suite of techniques to help train DNNs in a data efficient manner. We found that a simple CNN architecture, such as ResNet, is adequate for achieving good accuracy. This is in contrast to prior work employing complex architectures like hybrid CNN-RNN [1], non-local block additions to CNNs [12], etc.

In order to efficiently use the available data, we did extensive analysis of the ICBHI dataset. We found several characteristics of the data that might inhibit training DNNs effectively. For example, the dataset contains audio recordings from four different devices, with skewed distribution of samples across the devices, which makes it difficult for DNNs to generalize well across devices. Similarly, the dataset has a skewed distribution across normal and abnormal classes, and varying lengths of audio samples. We propose multiple novel techniques to address these problems—device specific fine-tuning, concatenation-based augmentation, blank region clipping, and smart padding. We perform extensive evaluation and ablation analysis of these techniques.

We evaluate the performance of our framework on the respiratory anomaly classification task proposed in the ICBHI challenge [21]. This is further divided into two subtasks: (i) classify a breathing cycle into one of the four classes–normal(N), crackle(C), wheeze(W), both(B), and (ii) classify a breathing cycle into normal or anomalous class, where $anomalous=\{crackle,wheeze,both\}$. Our evaluation method is same as the one proposed in the original ICBHI challenge. The final score is computed as the mean of Sensitivity $S_{e}=\frac{P_{c}+P_{w}+P_{b}}{N_{c}+N_{w}+N_{b}}$ and Specificity $S_{p}=\frac{P_{n}}{N_{n}}$, where $P_{i}$ and $N_{i}$ are the number of correctly classified and total number of samples in class $i$, respectively ($i\in\{normal,crackle,wheeze,both\}$). For the 2-class case, we adopt the anomalous and normal class scores as $S_{e}$ and $S_{p}$ respectively, and the score is computed as their mean.

We compare our performance using the above evaluation metric on two dataset divisions: the official 60-40$\%$ split [21] and the 80-20$\%$ split [11, 12, 1] for train-test⁴⁴4For both the splits, the train and test set are patient-wise disjoint.. The Sensitivity $S_{e}$, Specificity $S_{p}$ and ICBHI Score values are reported in Table 1. RespireNet achieves state-of-the-art (SOTA) in both train-test split divisions, and outperforms SOTA [12] on the official split (60-40) by $4\%$ and SOTA [1] on the 80-20 split by $2.2\%$. Further, RespireNet achieves a score of $77\%$ on the 2-class classification task, achieving the new SOTA.

We further analyze the effect of our novel proposed techniques by conducting an ablation analysis on the 4-class classification task on the 80/20 split.

Recently, there has been a lot of interest in using deep learning models for respiratory sounds classification [1, 12, 9]. It has outperformed statistical methods (HMM-GMM) [8] and traditional machine learning methods (boosted decision trees, SVM) [4, 24]. In these deep learning based approaches, a time-frequency representation of the audio signal is provided as input to the model. Kochetov et al. [9] propose a deep recurrent network with a noise masking intermediate step for the four class classification task, obtaining a score of $65.7\%$ on the 80-20 split. However the paper omits the details regarding noise label generation [1], thus making it hard to reproduce. Deep residual networks and optimized S-transform based features are used by Chen et al. [6] for three-class classification of anomalies in lung sounds. The model is trained and tested on a smaller subset of the ICBHI dataset on a 70-30 split and achieve a score of $98\%$.

Acharya and Basu [1] propose a Mel-spectrogram based hybrid CNN-RNN model with patient-specific model tuning, achieving a score of $66.3\%$ on 4-class and 80-20 split. Ma et al. [12] introduce LungRN+NL which incorporates a non-local block in the ResNet architecture and apply mixup augmentations to address the data imbalance problem and improve the model’s robustness, achieving sensitivity of $63.7\%$. However, none of these approaches focus on characteristics of the ICBHI dataset, which we exploit to improve performance.

The paper proposes RespireNet a simple CNN-based model, along with a set of novel techniques—device specific fine-tuning, concatenation-based augmentation, blank region clipping, and smart padding—enabling us to effectively utilize a small-sized dataset for accurate abnormality detection in lung sounds. Our proposed method achieved a new SOTA for the ICBHI dataset, on both the 2-class and 4-class classification tasks. Further, our proposed techniques are orthogonal to the choice of network architecture and should be easy to incorporate within other frameworks.

The current performance limit of the 4-class classification task can be mainly attributed to the small size of the ICBHI dataset, and the variation among the recording devices. Furthermore, there is lack of standardization in the 80-20 split and we found variance in the results based on the particular split. In future, we would recommend that the community should focus on capturing a larger dataset, while taking care of the issues raised in this paper.

This supplementary material includes some other details about the dataset, and additional results which could not be accommodated in the main paper.