Detection of breath cycles in pediatric lung sounds via an object detection-based transfer learning method

Abstract

Auscultation is critical for assessing the respiratory system in children; however, the lack of pediatric lung sound databases impedes the development of automated analysis tools. This study introduces an object detection-based transfer learning method to accurately predict breath cycles in pediatric lung sounds. We utilized a model based on the YOLOv1 architecture, initially pre-trained on an adult lung sound dataset (HF_Lung_v1) and subsequently fine-tuned on a pediatric dataset (SNUCH_Lung). The input feature was the log Mel spectrogram, which effectively captured the relevant frequency and temporal information. The pre-trained model achieved an F1 score of 0.900 ± 0.003 on the HF_Lung_v1 dataset. After fine-tuning, it reached an F1 score of 0.824 ± 0.009 on the SNUCH_Lung dataset, confirming the efficacy of transfer learning. This model surpassed the performance of a baseline model trained solely on the SNUCH_Lung dataset without transfer learning. We also explored the impact of segment length, width, and various audio feature extraction techniques; the optimal results were obtained with 15 s segments, a 2-second width, and the log Mel spectrogram. The model is promising for clinical applications, such as generating large-scale annotated datasets, visualizing and labeling individual breath cycles, and performing correlation analysis with physiological indicators. Future research will focus on expanding the pediatric lung sound database through auto-labeling techniques and integrating the model into stethoscopes for real-time analysis. This study highlights the potential of object detection-based transfer learning in enhancing the accuracy of breath cycle prediction in pediatric lung sounds and advancing pediatric respiratory sound analysis tools.