Respiratory sound classification for crackles, wheezes, and rhonchi in the clinical field using deep learning

Abstract

Auscultation has been essential part of the physical examination; this is non-invasive, real-time, and very informative. Detection of abnormal respiratory sounds with a stethoscope is important in diagnosing respiratory diseases and providing first aid. However, accurate interpretation of respiratory sounds requires clinician's considerable expertise, so trainees such as interns and residents sometimes misidentify respiratory sounds. To overcome such limitations, we tried to develop an automated classification of breath sounds. We utilized deep learning convolutional neural network (CNN) to categorize 1918 respiratory sounds (normal, crackles, wheezes, rhonchi) recorded in the clinical setting. We developed the predictive model for respiratory sound classification combining pretrained image feature extractor of series, respiratory sound, and CNN classifier. It detected abnormal sounds with an accuracy of 86.5% and the area under the ROC curve (AUC) of 0.93. It further classified abnormal lung sounds into crackles, wheezes, or rhonchi with an overall accuracy of 85.7% and a mean AUC of 0.92. On the other hand, as a result of respiratory sound classification by different groups showed varying degree in terms of accuracy; the overall accuracies were 60.3% for medical students, 53.4% for interns, 68.8% for residents, and 80.1% for fellows. Our deep learning-based classification would be able to complement the inaccuracies of clinicians' auscultation, and it may aid in the rapid diagnosis and appropriate treatment of respiratory diseases.

Conflict of interest statement

The authors declare no competing interests.

© 2021. The Author(s).

Figures

Figure 1

Scheme of the classification of respiratory sounds using deep learning. Lung sounds database contains normal sounds, crackles, wheezes, and rhonchi. Deep learning was used for two types of classification: The first step is the discriminating normal sounds from abnormal sounds. The second is to categorize abnormal sounds into crackles, wheezes, and rhonchi. (ER: Emergency room, ICU: intensive care unit).

Figure 2

ROC of the model for discrimination of abnormal lung sounds. Each plot illustrates the ROC of the algorithm on the independent testing set for abnormal lung sounds, with AUC of 0.93.

Figure 3

ROC of the model for classifying abnormal lung sounds into crackles, wheezes, and rhonchi. Each plot illustrates the ROC of the algorithm on the independent testing set for crackles, wheezes, and rhonchi with the mean AUC of 0.92.

Figure 4

Accuracy of auscultation analysis in real clinical practice. (A) Mean correction answer rates for the overall sounds, normal sounds, crackles, wheezes, and rhonchi. (B) Mean correction answer rates of students, interns, residents, and fellows for overall sounds. (C) Mean correction answer rates of students, interns, residents, and fellows for normal sounds (D) Mean correction answer rates of students, interns, residents, and fellows for crackles (E) Mean correction answer rates of students, interns, residents, and fellows for wheezes. (F) Mean correction answer rates of students, interns, residents, and fellows for rhonchi. *p < 0.05, **p < 0.05 ***p < 0.001 (Student's t-test).

Figure 5

Summary of deep learning assisted classification of respiratory sounds. Respiratory sounds were corrected from the patients with pulmonary diseases. The sounds were validated and classified by pulmonologists. The sounds were converted to Mel-spectrogram and features were extracted by VGG16 (transfer learning). Respiratory sounds were classified by CNN. Deep learning-based classification of respiratory sounds can be helpful for screening, monitoring, and diagnosis of pulmonary diseases.

Figure 6

Overview of our AI models.

Figure 7

Process to obtain spectrograms. (A) Given lung sound, dividing lung sound files with overlapping during 3 s (B) Obtaining three types of Mel-spectrograms with log-scale.

Figure 8

VGG16 architecture for our model. The input size for our model is 256 × 256. We freeze all layers in VGG16 without fully-connected layer, and with extracting features, we classified the respiratory sounds.

Figure 9

5-folds cross validation. Trying 5 iterations and getting final results with mean of all performances.