Domain Knowledge-Enhanced Integrated Model for LUS Video Scoring

Pneumonia, an acute respiratory infection marked by inflammatory responses typically manifesting in the alveoli, distal airways, and pulmonary interstitium, poses a substantial risk to patients' life and health due to its protean clinical presentations and rapid progression. Lung ultrasound videos, as a non-invasive and expeditious imaging modality, provide real-time monitoring of pulmonary lesions and other conditions. Assessing pulmonary status through lung ultrasound videos represents a practical and cost-effective diagnostic approach. We propose a domain knowledge-enhanced integrated model based on lung ultrasound videos. Firstly, we introduce the temporal-C3D network as the backbone of the model, which can extract spatial information of the videos while preserving temporal dynamics, aiding to better understand the temporal structure and dynamic changes of the videos. Subsequently, a clinical prior knowledge-based spatial attention module is used to weight the features of the designed temporal-C3D network. Furthermore, we add a temporal attention mechanism to complement its ability to capture temporal information. Through training and validation of collected ultrasound video clips, the model achieved good performance in the task of pneumonia status assessment, 87.76%, providing valuable reference for clinical diagnosis.