|
論文大綱︰ |
Respiratory diseases are a major health concern, highlighting the need for early detection. While electronic stethoscopes improve auscultation, diagnosis still heavily relies on clinical expertise. This study proposes a systematic framework for respiratory sound classification, spanning from model development to embedded deployment for real-time use. In the first stage, the OFGST-Swin method is introduced for respiratory cycle classification. It incorporates two novel modules: a Sliding Window-based Augmentation (SWA) technique to alleviate data imbalance by generating adventitious cycles, and the Overlap Fusion-based Generalized S-transform (OFGST) for detailed feature extraction. In the second stage, a lightweight model, LungNeXt, is proposed for respiratory sound classification. The lightweight design achieves efficient and accurate classification. The model integrates Rand Clip Mix (RCM) and Enhanced Mel-spectrogram Feature Extraction (EMFE). RCM augments the data by mixing sound clips within the same class, while EMFE enhances key frequency bands to improve feature quality and classification accuracy. In the third stage, the study presents LungScope, an embedded system deploying the LungLite model on a Raspberry Pi 4 Model B. The system includes a custom-designed expansion board and supports real-time acquisition, classification, and result display, enabling point-of-care diagnostics in low-resource settings. The systematic framework of this study improves the performance of respiratory sound classification. It also provides a practical solution for real time medical diagnosis in resource-limited environments. |
