Breathing Cycle Detection System for Respiratory Sounds Using Deep Learning

Computer-aided lung sound analysis (CALSA) paired with digital auscultation provides a cheap, objective, and non-invasive measurement option for healthcare providers. An important task of CALSA algorithms is the detection of the inspiration and expiration breathing phases. In this work, we propose a system that uses a deep learning model to predict the inspiration and expiration phases from a Mel-spectrograms of the lung sound. Multiple convolutional neural networks are investigated alongside vision transformer architectures. Due to the absence of standardized metrics for the evaluation of breathing cycle detection systems, we propose a novel intersection-over-union (IoU) score in order to provide a thorough evaluation. Our final system obtained a duration weighted mean IoU score of 0.82 and an F1 score of 80% for cycle-based detection. Moreover, we demonstrate that the performance of our system is on par with the cycle-labelling performance of respiratory experts and surpasses the performance of other comparable methods. The proposed system was trained, validated, and tested on diverse datasets with a respective 591 minutes, 66 minutes and 146 minutes of respiratory recordings.